Active promotion

A certain number of actors, both at the national and at the local level are pushing for the development and the extension of biometric remote identification. At the local level, figures such as Nice’s (France) mayor Christian Estrosi have repeatedly challenged Data Protection Authorities, arguing for the usefulness of such technologies in the face of insecurity (for a detailed analysis, see chapter 8 in this report, see also Barelli 2018). At the national level, Biometric systems for the purposes of authentication are increasingly deployed for forensic applications among law-enforcement agencies in the European Union. As we elaborate in Chapter 3, 11 out of 27 member states of the European Union are already using facial recognition against biometric databases for forensic purposes and 7 additional countries are expected to acquire such capabilities in the near future. Several states that have not yet adopted such technologies seem inclined to follow the trend, and push further. Former Belgian Minister of Interior Pieter De Crem for example, recently declared he was in favour of the use of facial recognition both for judicial inquiries but also for live facial recognition, a much rarer instance. Such outspoken advocates of the use of RBI constitute an important voice, but do not find an echo in the EU mainstream discussions.

Support with safeguards

While there is little widespread support for the development of centralised biometric mass surveillance, some actors such as the European Council and the European Commission have advocated a cautious and regulated development of remote biometric identification systems, as part of a broader Artificial Intelligence strategy. The principles of such strategies have been outlined in the various strategy documents discussed above. A large number of the technology companies are hoping that this position remains the main one, with many of them eager to implement the ethical requirements necessary for the deployments of their systems. In addition to the political and legislative activity mentioned above, the EU institutions have been active in promoting the use of Artificial Intelligence and biometric surveillance technologies. As detailed in chapter 5, instruments such as the Digital Europe programme, the Connecting Europe Facility 2 and Horizon Europe will form the basis for collaboration between public institutions and the security industry developing biometric remote identification products. In the European Parliament, positions are divided and moving, but parties like the European People Party support a similar notion of careful development.

Moratorium

For other actors, such as the European Parliament or the Council of Europe, remote biometric identification systems entail too many unknown risks and thus need to be put on hold. The proponents of a moratorium invoke the necessity of applying the principle of precaution – a similar strategy to opponents of the commercialisation of GMO in Europe – so that all dimensions of the technology can be assessed before a decision can be made. On 20 January 2021,  the European Parliament passed a resolution inviting the EU Commission to consider a moratorium on the use of facial recognition systems (European Parliament 2021). Similarly, in 2021, the Council of Europe (2021) adopted Guidelines on Facial Recognition (Council of Europe, 2021) which call for a moratorium for the live facial recognition technologies and lay out certain conditions for the use of facial recognition technologies by law enforcement authorities.)

Ban

Finally, a growing number of actors considers that there is enough information about remote biometric identification in public space to determine that they will never be able to comply to the strict requirement of the European Union in terms of respect of Fundamental Rights, and as such should be banned entirely. It is the current position of the European Data Protection Supervisor (EDPS, 2021) the Council of Europe and a large coalition of NGOs (among which La Quadrature du Net and the collaborative project Technopolice,) gathered under the umbrella of the European Digital Rights organisation (EDRi 2020). In the European Parliament, the position has most vocally been defended by the European Greens, but has been shared by several other voices, such as members of the Party of the European Left, the Party of European Socialists or Renew Europe (Breyer et al 2021).

Gait recognition

Gait recognition consists of recognising the specific way in which a person walks (gait), but in reality it covers a broader range of criteria (body, proportions, posture, etc.) (Segal 2020, 2). The advantages of gait recognition are that it does not require a clear access to a face, and it requires a lower image resolution (as it analyses an entire body, not only a face). Gait recognition, however, requires more computing power because it works on the basis of moving images (i.e., multiple frames of still images, up to 30 frames per second) rather than still images. Gait recognition has been used as evidence in court for a case in Denmark (Segal 2020, 18). Gait recognition poses important technical challenges: The amount of data storage and processing power far exceeds that of facial recognition. There are currently very few training datasets. So far, systems have proven to be more expensive, and less accurate than facial recognition.

People tracking and counting

This is perhaps the form of person tracking with which the least information about an individual is stored. An object detection algorithm estimates the presence and position of individuals on a camera image. These positions are stored or counted and used for further metrics. It is used to count passers-by in city centres, and for a one-and-a-half-meter social distancing monitor in Amsterdam³. See also the case study in this document on the Burglary-Free Neighbourhood in Rotterdam (CHAPTER 7), which goes into more detail about the use of the recorded trajectories of individuals to label anomalous behaviour.

Emotion recognition.

Software that categorises facial expressions into emotion categories – happiness, sadness, anger, etc. – is known to be used in billboards that are equipped with cameras, in order to analyse audience response to advertisements. For example, in airports or at train stations. While the face is claimed to be a “window into the brain” by some, the technology has been heavily criticised. Firstly, some consider it an undesirable invasion of their privacy, while other critique the technology for capturing primarily stereotypical ways of expressing oneself (van de Ven, 2017). In some places, such as at Dutch train stations, these critiques have led to disabling the cameras in billboards altogether (Het Parool, 2017).

Age, gender, and ethnicity classification

Aside from deducing emotions, the face is used to deduce a variety of traits from individuals. For example, gender, ethnicity, and age estimations are available in many off-the-shelf facial analysis products. As with emotion recognition, these classifications are mainly used in digital signage and video advertisement contexts. LGBTQ+ communities have spoken out against automatic gender classification, pointing out that a long fought, non-binary understanding of gender is made undone by the technology’s binary classifications (Vincent, 2021). Similarly, recent revelations that Hikvision (China) has used similar technology to estimate whether an individual is from China’s Uyghur minority, has directly led the European Parliament to call for a ban of Hikvision’s products on the Parliament’s premises (Rollet, 2021).

Audio recognition

From a technological perspective, neural networks process audio relatively similarly to how video is processed: rather than feeding an image, a spectrogram is used as input for the network. However, under the GDPR, recording conversations, is illegal in the European Union without informed consent of the participants. In order to adhere to these regulations, on some occasions, only particular frequencies are recorded and processed. For example, in the Burglary-Free Neighbourhood in Rotterdam (Netherlands) (CHAPTER 7), only two frequencies are used to classify audio; making conversations indiscernible while being able to discern shouting or the breaking of glass⁴. Another initiative using audio in to enhance the surveillance camera is the Living Lab International Zone project in the Hague (Netherlands), a collaboration between a broad range of partners⁵.

Image acquisition: Controlled and uncontrolled images

Facial recognition begins with an image. An image which will be subject to the algorithm’s scrutiny. Controlled images are images that are captured for the purpose of processing, aimed at optimal positions and lighting conditions. They are for example taken at a police station, or at a photographer’s studio with strict requirements, and are either contained in databases that precede the introduction of a facial recognition system (e.g., driver’s license databases) or are specifically designed to match high criteria of biometric systems (i.e., photographs for biometric passports). Uncontrolled images are images that are captured outside of specific requirement, collected for example through social media scraping or video surveillance.

When it comes to the acquisition technologies (cameras) for uncontrolled images, over the past decades, the main evolution in terms of video has been the passage from analogue video to digital video, the latter allowing images to be processed through computers. As in the realm of consumer cameras, the initial race was for better definition (calculated in terms of megapixels). “Smart” camera systems require a slightly higher resolution than standard video surveillance systems in order to guarantee a minimum of 300 PPM to adequately feed the software (IPVM Team 2020, 5). But overall, the average camera does not exceed a definition of 4 megapixels and are more often in the area of 2 megapixels (which yields a 1080p or HD resolution)⁶. The quality of capture, especially in non-cooperative scenarios, is determined by two main external variables: the angle of the face relative to the camera (front, side, back, top) and the lighting conditions (bright daylight, dark night). In recent years, manufacturers have added an additional infra-red channel to the red-green-blue (RGB) video channels in order to increase detail accuracy in low-light conditions.

What makes systems “smart”: image processing algorithms

The processing of the photographic or video image by a specific software application is where the “smart” processing happens. Broadly speaking video surveillance technology can be split in two key historical moments: before machine learning, and after machine learning.

Video motion detection (VMD) and heuristic filters. The early smart technologies relied on simple motion detection algorithms which compared pixel changes from one image to the next (Quevillon 2012). The problem is that any movement (the leaves of a tree) or change of light (a car passing in the night) can trigger the systems. Heuristic filters were thus added to VMD systems in order to give additional parameters to the system (amount and size of pixel changing etc.). Both systems were highly inefficient and prone to trigger false alarms, making such technologies unattractive. The main problem was that only pre-established changes hard coded by humans would be detected by the systems.

Machine learning. Machine learning revolutionised image-based biometric identification. Machine learning is an automated process through which the software application will be programmed to recognise particular patterns, based on a dataset it is “trained” on. There are three ways in which this configuration of the machine learning model can be controlled: supervised, semi-supervised or unsupervised. Supervised machine learning consists of teaching the system to recognise people, cars, guns, or any other object by feeding it an annotated dataset of such objects. It is supervised because humans “supervise” how the computer learns, by annotating the dataset (“this is a car”, “this is a gun” etc.). The categories of the annotations (cars, guns, etc.) will thus be the only ones that the system will be able to recognise (if only cars and guns are annotated, the system won’t in such a case recognise cats). Most video surveillance systems use supervised machine learning (IPVM Team 2021a, 11). Unsupervised machine learning lets the system cluster objects by itself. The advantage is the open-endedness of the systems (meaning they can generate categories of objects not anticipated in the training dataset), but the disadvantage is that algorithms can potentially cluster objects along irrelevant criteria for the task (for example clustering red motorcycles, cars, and trucks in one group and green ones in another, as opposed to creating one cluster for all motorcycles, one for cars and one for trucks). For this reason, semi-supervised machine learning, where only a small part of the data is labelled, can be used. Currently not widely in use, unsupervised machine learning is a growing trend in the video surveillance sector (IPVM Team 2021a, 12–13).

Both supervised and unsupervised learning exist in many shapes and sizes. For example, the Viola-Jones object detection algorithm⁷ from 2001, which made real-time face detection viable, is a supervised algorithm. Contemporary developments in video processing focus on using various kinds of artificial neural networks (i.e., convolutional neural networks, recurrent neural networks) to classify images and videos. These networks can be trained either supervised, semi-supervised or unsupervised depending on their configuration.

Machine learning and operational datasets

Remote biometric identification and classification relies in large part on datasets, for two key but distinct moments of their operation.

Machine learning datasets. These are the datasets used to train models through machine learning. We find three categories of such datasets. Publicly available datasets for object detection such as COCO, ImageNet, Pascal VOC include a varying number of images labelled in a range of categories, these can be used to train algorithms to detect for example people on an image (IPVM Team 2021a, 27). The most used open-source datasets for surveillance technologies are Celeb 500k, MS-Celeb-1Million-Cleaned, Labeled Faces in the Wild, VGG Face 2, DeepGlint Asian Celeb, IMDB-Face, IMDB-Wiki, CelebA, Diveface, Flickr faces and the IARPA Janus Benchmark (IPVM Team 2021b, 7). Many of these datasets also function as a public benchmark, against which the performance and accuracy of various algorithms is measured. For example, Labeled Faces in the Wild, the COCO dataset and NIST present such leaderboards on their website⁸. Government datasets are generally collections of images available to a government for other purposes (driver’s license, passport, or criminal record photo datasets). While in Europe most of these datasets are not accessible to the public, in China and in the US, they are made available for testing and training purposes to private companies, such as the Multiple Encounter Dataset (NIST, 2010). Finally proprietary datasets may be developed by providers for their specific applications.

Machine learning models. In the machine learning process, an algorithm gets iteratively configured for the optimal output, based on the particular dataset that it is fed with. This can be a neural network, but also e.g., the aforementioned Viola-Jones’ object detector algorithm. The model is the final configuration of this learning process. As such, it does not contain the images of the dataset in and of themselves. Rather, it represents the abstractions the algorithm “learned” over time. In other words, the model operationalises the machine learning dataset. For example, the YOLO object detection algorithm yields different results when it is trained on either the COCO or the model (in conjunction with the algorithm) which determines the translation of an image into a category, or of the image of a face into its embedding.

Operational datasets, or image databases. Datasets used in training machine learning models should be distinguished from matching or operational datasets which are the “watchlists” of for example criminals, persons of interest or other lists of individuals against which facial recognition searches will be performed – whether these are in real time or post hoc. These datasets contain pre-processed images of individuals on the watchlist, and store the numerical representations of these faces, their feature vectors or embedding, in an index for fast retrieval and comparison with the queried features (using for example k-Nearest Neighbour or Support Vector Machines). Face or object detection models do not use such a dataset.

Availability

Facial recognition algorithms can be developed in-house, taken from an open-source repository, or purchased (IPVM Team 2021b, 14). Popular open-source facial recognition implementations include OpenCV, Face_pytorch, OpenFace and Insightface. Many of these software libraries are developed at universities or implement algorithms and neural network architectures presented in academic papers. They are free, and allow for a great detail of customisation, but require substantial programming skills to be implemented in a surveillance system. Moreover, when using such software, the algorithms run on one’s own hardware which provides the developer with more control, but also requires more maintenance.

Proprietary facial recognition. There are three possible routes for the use of proprietary systems: There are “turnkey” systems sold by manufacturers such as Hikvision, Dahua, AnyVision or Briefcam. Those integrate the software and hardware, and as such can be directly deployed by the client. Algorithm developers such as Amazon AWS Rekognition (USA), NEC (Japan), NTechlab (Russia), Paravision (USA) allow to implement their algorithms and customise them to one’s needs, and finally there are “cloud” API systems, a sub-set of the former category, where the algorithm is hosted in a datacentre and is accessed remotely (IPVM Team 2021b, 16). The latter type of technology bears important legal ramifications, as the data may travel outside of national or European jurisdictions. It should be noted that many of the proprietary products are based on similar algorithms and network architectures as their open-source counterparts (OpenCV, 2021). Contrary to the open-source software, it is generally unclear which datasets of images have been used to train the proprietary algorithms.

Data capture challenges

Facial recognition’s accuracy can easily be challenged by a certain number of factors in the capture of the data to be analysed, in particular when dealing with “non-cooperative” image capture. The resolution of the camera, and in particular the key variable of Pixels per Meter (minimum 300 PPM is generally required) is instrumental in ensuring that enough information is provided to the algorithm. Lighting conditions are similarly important. Although increasingly cameras add an infra-red channel to the RGB channels in order to recover detail in low-light conditions, inadequately illuminated faces will generate a high number of errors. Orientation of the face in relation to the camera is one more key factor to take into account, especially because a camera will rarely be mounted at face level (more likely overhead), and thus difficult angles will often result in partial representation of faces (Fernandez et al. 2020, 29). Vision can often be blocked by other factors, such as other individuals in large crowds, sunglasses, masks (in particular in times of COVID-19). Obstruction can be voluntary when individuals for example look down to avoid surveillance. Finally, not all systems have a liveness detection system, meaning that they can be tricked by a photograph of a face instead of a real face. (IPVM Team 2020, 12–13)

Dataset-related challenges

Datasets also face a number of technical challenges. For machine learning systems, small datasets will inadequately train the algorithms, simply because there are not enough different instances of the type of face or object that is supposed to be recognised. This is a challenge for gait recognition algorithms for example, for which there is a dearth of large datasets. Changes in features (such as hair, facial hair, beard, earrings) in the dataset can lead to a poorly trained algorithm. Datasets are often labelled with a specific purpose, and thus training an algorithm on a dataset that is not representative of the use-case can provide counter-productive results.

More problematically, a lack of diversity, in particular when it comes to ethnicity, age, or gender leads to bias in the algorithm. This issue has been at the core of the US-based discussion on the banning of Facial Recognition. Public databases such as VGGFace2 (based on faces from Google images) and MS-Celeb-1M42 (celebrity faces) are often used to train facial recognition algorithms yet are far from representative of everyday populations – this is called representation bias (Fernandez et al. 2020, 30). The main goal of the project Gender Shades led by Joy Buolamwini was both to show the lack of representativity of existing datasets and address the problem of the consequent discrepancy between the error rates related to light-skinned men and dark-skinned women (Fernandez et al. 2020, 30–31).

However, a representational dataset is not always a desirable dataset, because actual structural biases often do not match the values of society. Illustrative of this is that, when doing a Google image search for the term “CEO” it would originally return primarily photographs of white male people. While this was representative of the CEO population (and thus accurate), the results reinforce the vision of a world that does not align with progressive societal values (Suresh, 2019). Because of the gap between ideals of equality and actual societal structural inequalities, datasets can be either representative of an unequal society, or representative of desired equality – but never of both at the same time.

Datasets upon which the computer algorithm will later be able to distinguish particular entities or behaviour are built through vast amounts of human labour. For example, the work that has gone into the image dataset ImageNet is equivalent to 19 years of working 24 hours a day, 7 days a week (Malevé, 2020). Nevertheless, quantity does not necessarily equal quality. Many of the categories with which images are annotated are ambiguous. Not in their dictionary definition per se, but when they enter the culture of the annotation workers. For example, the category of “ratatouille” contains images of various stews, salads and even a character of the eponymous Pixar movie. Similarly, the category “Parisian” contains images of Paris Hilton (Malevé, 2020). This ambiguity of categories does not only haunt ImageNet. The aforementioned COCO dataset contains images of a birdhouse in the shape of a bird, which is tagged as bird, or a bare pizza bottom which is tagged as pizza (Cochior and van de Ven, 2020). These examples show that even seemingly unambiguous concepts become fluid the moment they have to become strictly delineated in a dataset.

Another important issue with ethical and political repercussions is unethically collected data, as in the case of Clearview AI detailed above. When it comes to operational datasets, i.e., datasets used in the actual process of facial authentication and/or identification, we have seen that possible deployments include the use of cloud-based services (either for the processing or the storage of the sensitive information). This increases the risks of data breaches and attacks by hackers. (Fernandez et al. 2020, 34)

Algorithm-related challenges

Finally, there are issues related to the quality and performance of the algorithms and how to measure it. The National Institute of Standards and Technology is an agency of the US Department of Commerce. The NIST provides the possibility for vendors to test the efficacy of their algorithms on a standardised dataset, the “Ongoing Face Recognition Vendor Test (FRVT).

As an IPVM study shows, brands often use single-number scores obtained from NIST vendor tests (i.e., “our algorithm showed 98,6% accuracy”.) (IPVM Team 2021b, 17). These scores are however obtained in very controlled conditions that do not match the real-world use of the algorithms. There are thus important discrepancies in this regard. Moreover, the accuracy score is not always representative of desirable behaviour of a model. Data scientists therefore distinguish precision and recall, to better account for cases where e.g., positive classification is rare, yet of high impact – for example when classifying individuals as high risk (Shung 2020, 202). These distinctions are often lost in the commercial language and in the public debate.

A final issue related to working with the existing algorithms is what is known as observer bias or confirmation bias. The output of an algorithm reinforces the (subconscious) biases that went into producing it. It can occur both when creating the dataset or when training and running the algorithms. For example, the software used for predictive policing in Chicago helped determine where to send police officers on patrol. “Because these predictions are likely to over‐represent areas that were already known to police, officers become increasingly likely to patrol these same areas and observe new criminal acts that confirm their prior beliefs regarding the distributions of criminal activity. The newly observed criminal acts that police document as a result of these targeted patrols then feed into the predictive policing algorithm on subsequent days, generating increasingly biased predictions. This creates a feedback loop where the model becomes increasingly confident that the locations most likely to experience further criminal activity are exactly the locations, they had previously believed to be high in crime.” (Lum and Isaac, 2016). The example reveals that the different kinds of biases at play are hard to untangle, as the observer bias coincides with a historical bias of over-policing. It requires a lot of work to recognise such confirmation biases in the automated operation of automated classification software. The “black box” dimension of their operation – and the only just emerging efforts to build explanatory AI – make it difficult to understand their categorisation process (Xie et al. 2020; Fernandez et al. 2020, 34)

Live authentication

As in the cases of the use of Cisco systems powered FRT in two pilot projects in high schools of Nice (see section 8.1) and Marseille (France)⁹, or as in the case of the Anderstorp Upper Secondary School in Skelleftea (Sweden)¹⁰, the aim of these projects was to identify students who could have access to the premises. School-wide biometric databases were generated and populated with students’ portraits. Gates were fitted with cameras connected to facial recognition technology and allowed access only to recognised students. Another documented use has been for the Home Quarantine App (Hungary), in which telephone cameras are used by authorities to verify the identity of the persons logged into the app (see also section 10.1).

In these deployments, people must submit themselves to the camera in order to be identified and gain access. While these techniques of identification pose important threats to the privacy of the concerned small groups of users (in both high school cases, DPAs banned the use of FRTs), and run the risk of false positives (unauthorised people recognised as authorised) or false negatives (authorised people not recognised as such) the risk of biometric mass surveillance strictly speaking is low to non-existent because of the nature of the acquisition of images and other sensor-based data.

However, other forms of live authentication tie in with surveillance practices, in particular various forms of blacklisting. With blacklisting the face of every passer-by is compared to a list of faces of individuals who have been rejected access to the premises. In such an instance, people do not have to be identified, as long as an image of their face is provided. This has been used in public places, for example in the case of the Korte Putstraat in the Dutch city of 's-Hertogenbosch: during the carnival festivities of 2019 two people were rejected access to the street after they were singled out by the system (Gotink, 2019). It is unclear how many false positives were generated during this period. Other cases of blacklisting can be found at, for example, access control at various football stadiums in Europe, see also section 3.3. In many cases of blacklisting, individuals do not enrol voluntarily.

Forensic authentication

Biometric systems for the purposes of authentication are also increasingly deployed for forensic applications among law-enforcement agencies in the European Union. The typical scenario for the use of such technologies is to match the photograph of a suspect (extracted, for example, from previous records or from CCTV footage) against an existing dataset of known individuals (e.g., a national biometric database, a driver’s license database, etc.). (TELEFI, 2021). The development of these forensic authentication capabilities is particularly relevant to this study, because it entails making large databases ready for searches on the basis of biometric information.

To date, 11 out of 27 member states of the European Union are using facial recognition against biometric databases for forensic purposes: Austria (EDE)¹¹, Finland (KASTU)¹², France (TAJ)¹³, Germany (INPOL)¹⁴, Greece (Mugshot Database)¹⁵, Hungary (Facial Image Registry)¹⁶, Italy (AFIS)¹⁷, Latvia (BDAS)¹⁸, Lithuania (HDR)¹⁹, Netherlands (CATCH)²⁰ and Slovenia (Record of Photographed Persons)²¹ (TELEFI 2021).

Seven additional countries are expected to acquire such capabilities in the near future: Croatia (ABIS)²², Czech Republic (CBIS)²³, Portugal (AFIS) Romania (NBIS)²⁴, Spain (ABIS), Sweden (National Mugshot Database), Cyprus (ISIS Faces)²⁵, Estonia (ABIS)²⁶ (TELEFI 2021).

When it comes to international institutions, Interpol (2020) has a facial recognition system (IFRS)²⁷, based on facial images received from more than 160 countries. Europol has two sub-units which use the facial recognition search tool and database known as FACE: the European Counter Terrorism Center (ECTC) and the European Cybercrime Center (ECC). (TELEFI, 2021 149-153) (Europol 2020)

Only 9 countries in the EU so far have rejected or do not plan to implement FRT for forensic purposes: Belgium (see CHAPTER 6), Bulgaria, Denmark, Ireland, Luxembourg, Malta, Poland, Portugal, Slovakia.

Figure 1. EU Countries use of FRT for forensic applications²⁸

When it comes to databases, some countries limit the searches to criminal databases (Austria, Germany, France, Italy, Greece, Slovenia, Lithuania, UK), while other countries open the searches to civil databases (Finland, Netherlands, Latvia, Hungary).

This means that the person categories can vary substantially. In the case of criminal databases it can range from suspects and convicts, to asylum seekers, aliens, unidentified persons, immigrants, visa applicants. When civil databases are used as well, such as in Hungary, the database contains a broad range of “individuals of known identity from various document/civil proceedings” (TELEFI 2021, appendix 3).

Finally, the database sizes, in comparison to the authentication databases mentioned in the previous section, are of a different magnitude. The databases of school students in France and Sweden, mentioned in the previous section contains a few hundred entries. National databases can contain instead several millions. Criminal databases such as Germany’s INPOL contains 6,2 million individuals, France’s TAJ 21 million individuals and Italy’s AFIS 9 million individuals. Civil databases, such as Hungary’s Facial Image Registry contain 30 million templates (TELEFI, 2021 appendix 3).

Authentication has also been deployed as part of integrated “safe city” solutions, such as the NEC Technology Bio-IDiom system in Lisbon and London, deployed for forensic investigation purposes. For this specific product, authentication can occur via facial recognition, as well as other biometric authentication techniques such as ear acoustics, iris, voice, fingerprint, and finger vein recognition. We currently do not have public information on the use of Bio-IDiom in Lisbon nor in London. On NEC’s Website (2021) however, Bio-IDiom is advertised as a “multimodal” identification system, that has been used for example by the Los Angeles County Sheriff’s Department (LASD) for criminal investigations. The system “combines multiple biometric technologies including fingerprint, palm print, face, and iris recognition” and works “based on the few clues left behind at crime scenes. In Los Angeles, “this system is also connected to the databases of federal and state law enforcement agencies such as the California Department of Justice and FBI, making it the world’s largest-scale service-based biometrics system for criminal investigation”. We don’t know if that is the case in Portugal and in the UK deployments.

Case study: INPOL (Germany)

In order to give a concrete example of the forensic use of biometric technology, we can take the German case. Germany has been using automated facial recognition technologies to identify criminal activity since 2008 using a central criminal information system called INPOL (Informationssystem Polizei), maintained by the Bundeskriminalamt (BKA), which is the federal criminal police office. INPOL uses Oracle Software and includes the following information: name, aliases, date and place of birth, nationality, fingerprints, mugshots, appearance, information about criminal histories such as prison sentences or violence of an individual, and DNA information. However, DNA information is not automatically recorded (TELEFI 2021).

The INPOL database includes facial images of suspects, arrestees, missing persons, and convicted individuals. For the purpose of facial recognition, anatomical features of a person's face or head as seen on video surveillance or images are used as a material to match with data in INPOL. The facial recognition system compares templates and lists all the matches ordered by degree of accordance. The BKA has specific personnel visually analysing the system's choices and providing an assessment, defining the probability of identifying a person. This assessment can be used in a court of law if necessary (Bundeskriminalamt, n.d.). Searches in the database are conducted by using Cognitec Face VACS software (TELEFI 2021).

As of March 2020, INPOL consists of 5,8 million images of about 3,6 million individuals. All police stations in Germany have access to this database. The BKA saves biometric data and can be used by other ministries as well, for instance, to identify asylum seekers. Furthermore, the data is shared in the context of the Prüm cooperation on an international level (mostly fingerprints and DNA patterns). Furthermore, the BKA saves DNA analysis data as part of INPOL, accessible for all police stations in Germany. That database contains 1,2 million data sets (Bundeskriminalamt, n.d.). Other recorded facial images, for instance, driver’s licenses or passports, are not included in the search, and the database is mainly used for police work (TELEFI 2021).

A blurred boundary between authentication and surveillance

In principle, because of the strict legal framework to which law enforcement agencies are submitted, forensic biometric identification should not present a risk of biometric mass surveillance. The acquisition of images and the subsequent one-to-one searches are carried out as part of judicial investigations when a legal threshold of suspicion is met. The operation of the system by specialised forensic departments should follow the procedural limits set by the judicial process.

Function creep is however particularly concerning. If the legal framework is altered to allow the acquisition of live video, and if live searches on these are legally authorised against existing criminal and civil databases, then from a technical perspective it can be argued that there is potentially a risk of biometric mass surveillance. The main risk here being that the individuals whose identities are searched or tagged are not selected as a result of a judicial investigation, but indiscriminately. The system in place would then allow for search of these individuals against huge databases. In other words, by creating new biometric databases or upgrading existing databases to be FRT-readable, and developing or acquiring algorithmic capabilities to search them, law enforcement agencies across Europe are building an infrastructure which is technically capable of “switching” easily to a mode of operation akin to biometric mass surveillance.

Smart surveillance features

A first range of deployments of “smart” systems correspond to what can broadly be defined as “smart surveillance” yet do not collect or process biometric information per se²⁹. Smart systems can be used ex-post, to assist CCTV camera operators in processing large amounts of recorded information, or can guide their attention when they have to monitor a large number of live video feeds simultaneously. Smart surveillance uses the following features:

- Anomaly detection. In Toulouse (France), the City Council commissioned IBM to connect 30 video surveillance cameras to software able to "assist human decisions" by raising alerts when "abnormal events are detected." (Technopolice 2021) The request was justified by the “difficulties of processing the images generated daily by the 350 cameras and kept for 30 days (more than 10,000 images per second)”. The objective, according to the digital direction is "to optimise and structure the supervision of video surveillance operators by generating alerts through a system of intelligent analysis that facilitates the identification of anomalies detected, whether: movements of crowds, isolated luggage, crossing virtual barriers north of the Garonne, precipitous movement, research of shapes and colour. All these detections are done in real time or delayed (Technopolice 2021). In other words, the anomaly detection is a way to operationalise the numerical output of various computer vision based recognition systems. Similar systems are used in the Smart video surveillance deployment in Valenciennes (France) or in the Urban Surveillance Centre (Marseille).

- Object Detection. In Amsterdam, around the Johan Cruijff ArenA (Stadium), the city has been experimenting with a Digitale Perimeter (digital perimeter) surveillance system. In addition to the usual features of facial recognition, and crowd monitorining, the system includes the possibility of automatically detecting specific objects such as weapons, fireworks or drones. Similar features are found in Inwebit’s Smart Security Platform (SSP) in Poland.

- Feature search. In Marbella (Spain), Avigilon deployed a smart camera system aimed at providing “smart” functionalities without biometric data. Since regional law bans facial and biometric identification without consent, the software uses “appearance search”. “Appearance search” provides estimates for “unique facial traits, the colour of a person’s clothes, age, shape, gender and hair colour”. This information is not considered biometric. The individual’s features can be used to search for suspects fitting a particular profile. Similar technology has been deployed in Kortrijk (Belgium), which provides search parameters for people, vehicles and animals (Verbeke 2019). During the Covid-19 pandemic, several initiatives emerged to automatically detect whether the mask mandates were observed by the public, such as in the aborted face mask recognition project in Châtelet-Les Halles developed by the company Datakalab.

- Video summary. Some companies, such as Briefcam and their product Briefcam Review, offer a related product, which promises to shorten the analysis of long hours of CCTV footage, by identifying specific topics of interest (children, women, lighting changes) and making the footage searchable. The product combines face recognition, license plate recognition, and more mundane video analysis features such as the possibility to overlay selected scenes, thus highlighting recurrent points of activity in the image. Briefcam is deployed in several cities across Europe, including Vannes, Roubaix (in partnership with Eiffage managed by the City of Roubaix and the Métropole Européenne de Lille) and Moirans in France (with equipment provided by Nomadys).

- Object detection and object tracking. As outlined in chapter 2, object detection is often the first step in the various digital detection applications for images. An ‘object’ here can mean anything the computer is conditioned to search for: a suitcase, a vehicle, but also a person; while some products further process the detected object to estimate particular features, such as the colour of a vehicle, the age of a person. However, on some occasions — often to address concerns over privacy — only the position of the object on the image is stored. This is for example the case with the test of the One-and-a-half-meter monitor in Amsterdam (Netherlands), Intemo’s people counting system in Nijmegen (Netherlands), the ViSense social distancing monitor at MINDBase, a testing location of the Dutch Defence Equipment Organization; the KICK project in Brugge, Kortrijk, Ieper, Roeselare and Oostende (Belgium), the ViSense project in Mechelen (Belgium) or the Eco-counter tracking cameras pilot project in Lannion (France).

- Movement recognition. Avigilon’s software that is deployed in Marbella (Spain) also detects unusual movement. “To avoid graffiti, we can calculate the time someone takes to pass a shop window, “explained Javier Martín, local chief of police in Marbella to the Spanish newspaper El País. “If it takes them more than 10 seconds, the camera is activated to see if they are graffitiing. So far, it hasn’t been activated.” (Colomé 2019) Similar movement recognition technology is used in, the ViSense deployment at the Olympic Park London (UK) and the security camera system in Mechelen-Willebroek (Belgium). It should be noted that movement recognition can be done in two ways: where projects such as the Data-lab Burglary-free Neighbourhood in Rotterdam (Netherlands)³⁰ are only based on the tracking of trajectories of people through an image (see also ‘Object detection’), cases such as the Living Lab Stratumseind³¹ in Eindhoven (Netherlands) also process the movements and gestures of individuals in order to estimate their behaviour.

Integrated solutions

Deployment of RBI in public spaces

Here are the documented cases of RBI in public spaces we could find through our research:

- Live Facial Recognition pilot project in Brussels International Airport / Zaventem (Belgium, see detailed case study, CHAPTER 6)

- Live Facial Recognition in Budapest (Hungary, see detailed case study, CHAPTER 10)

- Live Facial Recognition pilot project during the Carnival in Nice (France, see detailed case study, CHAPTER 8)

- Live Facial Recognition Pilot Project Südkreuz Berlin (Germany, see detailed case study, CHAPTER 9)

As most of these cases are extensively discussed in the following chapters, we do not comment further on them here.

Additional cases are the Live Facial Recognition pilot during Carnival 2019 in 's-Hertogenbosch’s Korte Putstraat (the Netherlands) and the pilot of Live Facial Recognition in the city of Como³³, recently struck down by the Italian DPA (Garante per la Privacy). The deployment of facial recognition in Estacion Sur in Madrid (Spain) is also live.

Deployment of RBI in commercial spaces

The number of deployments of live facial recognition systems in commercial spaces hosting the public is much higher, but because of its commercial nature, difficult to document and trace. Our research found the following instances:

- Live Facial Recognition project, Brøndby IF Football stadium (Denmark)

- Live Facial Recognition Pilot in Metz Stadium (France)

- Live Facial Recognition in Ifema (Spain)

- Live Facial Recognition in Mercadona or Mallorca, Zaragoza, Valencia (Spain)

The systems operate more or less in the same way as RBI in public spaces, or as forensic authentication systems if they were connected to live cameras. In the Brøndby IF Football stadium deployment for example, developed in partnership with Panasonic and the National University of Singapore, the football fans who want to access the game have to pass through a gate equipped with a camera, connected to a facial recognition algorithm. The stadium administration has constituted a database of unwanted individuals and if the software matches one of the incoming fans with a record in the database, it flags it to the system (Overgaard 2019).

There is however little to no information of the uses of these technologies in commercial spaces, as there is no requirement to publicise the various components of these systems. The case studies of this report thus focus mostly on the deployment of RBI in public spaces. More research, and more transparency would however be welcome in order to understand the data gathering practices and the impact of these deployments.

The scope of the fundamental right to protection for RBI 

Article 7 of the EU Charter of Fundamental Rights (Charter) sets out national and EU legislators’ obligations on guaranteeing the right to private life, family life, and communications of individuals (the right to privacy) under EU law. The right to privacy can also be found in Article 8 of the European Convention on Human Rights (ECHR), the scope of which has evolved over the years to cover issues relating to the processing of personal data. Because Article 7 of the Charter mirrors closely Article 8 ECHR, its scope must be interpreted in line with the latter and its interpretation by the European Court of Human Rights (ECtHR) pursuant to Article 52(3) of the Charter. The Charter enshrines a separate right to protection of personal data in its Article 8, which is “distinct from Article 7 of the Charter” (C-203/15, Tele2, para. 129). 

Biometric surveillance tools interfere with the fundamental rights to privacy and personal data protection as enshrined in each of these legal sources because they collect, retain, process and use personal data, including an intrinsically special category of biometric data, which is - as discussed below, personal data relating to the physical, physiological or behavioural characteristics of an individual that allows their unique identification (see section 4.2.1). Notably, it may not be just the physical biometric data such as fingerprints (S and Marper v UK; C-291/12 , Schwarz) or facial images (Gaughran v UK) that benefits from the rights to privacy and personal data protection as enshrined in the ECHR and EU law. For example, the ECtHR has adopted an expansive approach in terms of recognising the protective scope of Article 8 ECHR (S and Marper v UK, para 67), which would afford protection to different categories of biometric data including behavioural biometric data such as one’s way of movement or voice (Venier and Mordini, 2010).

Privacy and data protection in public space and the risk of mass surveillance

The use of a wide range of biometric data discussed above engages with the individuals’ right to privacy and data protection even if they are captured and used in public spaces while individuals enjoy public life. The case law of the ECtHR (PG and JH v UK; Peck v UK) and the Court of Justice of the European Union (CJEU) (Opinion 1/15) shows that they have afforded privacy protection to information that is not inherently private. In fact, performing biometric surveillance in public spaces is inherently intrusive and amounts to mass surveillance, which in this context can simply be characterised as monitoring, tracking, or processing of personal data of individuals indiscriminately and in a generalised manner without a prior criminal suspicion (FRA 2018). Biometric surveillance in public spaces relies on generalised and indiscriminate collection, retention, use and sharing of biometric data of individuals. This is the case even if the intended purpose of the biometric surveillance is targeted, because in order to identify people on the watchlist in a crowd, every person in that particular space must be analysed and compared with the watchlist (Houwing 2020).

The grave consequences of this type of indiscriminate and generalised collection of personal data on fundamental rights of individuals can be found across the case law of the ECtHR and the CJEU. The ECtHR has repeatedly warned that covert surveillance tools must not be used to undermine or even destroy democracy on the grounds of defending it (Klass and others v Germany, para 49). Particularly in considering the lawfulness of collection of biometric data, the ECtHR recognised in S and Marper v UK that the use of biometric data that would allow identification of an individual and would carry the potential to deduce personal data that is classified as sensitive data such as ethnic origin would make the people concerned fundamentally vulnerable to stigmatisation and discrimination (paras 122-126). Because of the heightened level of protection afforded to it, the ECtHR found that generalised and indiscriminate collection and retention of biometric data did not comply with the ECHR requirements as it amounted to disproportionate interference with the right to privacy and thus constitute a violation of Article 8 ECHR. 

The CJEU considered in Digital Rights Ireland (Joined Cases C293/12 and C594/12, para 37) as well as Tele2 (C-203/15, para 100) that EU law precluded the mass retention of traffic and location data for law enforcement purposes, and only allowed for targeted retention of said data. The deployment of biometric surveillance in public spaces thus must be subject to strict scrutiny and in light of the case law of both courts, the EU fundamental rights law as well as the ECHR preclude the deployment of biometric surveillance that leads to mass surveillance for law enforcement purposes in public spaces. 

The ambiguities of “targeted” biometric surveillance

Targeted biometric surveillance may still be lawful provided that it is justified under Article 52(1) of the Charter in light of the ECHR requirements for the Convention rights that are mirrored in the Charter. This type of surveillance is distinguishable from mass surveillance as it is directed towards a person or group of persons based on a prior suspicion on their involvement with criminal activities. Recently in its La Quadrature du net and others decision, the CJEU added a geographical criterion as a satisfactory limitation for a targeted retention of traffic and location data (para 149). However, in the context of conducting biometric surveillance in public spaces, this might not be a limitation as such. As mentioned above, by its nature this type of surveillance would amount to mass surveillance since it would indiscriminately monitor and analyse everyone in that space to detect people on the watchlist. Accordingly, using biometric surveillance in a specific area (e.g., concert venues, football stadiums, public rallies) for law enforcement purposes might be considered as expansive and intrusive to an extent that it would constitute disproportionate interference with the rights to privacy and personal data protection.

Conditions for “targeted” biometric surveillance

Even where biometric surveillance is performed in a targeted way, its lawfulness would turn on the legitimate aim for which it is conducted and an assessment on its proportionality in light of that aim. A key issue here is that conducting targeted biometric surveillance in public spaces would constitute a serious interference with the rights to privacy and personal data protection because of the special character of biometric data that makes a person unique and identifiable and potentially carries the risk to reveal sensitive data. Thus, it should be conducted for an aim that is proportionate to the level of intrusiveness caused by it (by analogy C-203/15, Tele2, para 102). In essence, this means that targeted biometric surveillance is only allowed if it is strictly necessary for the purpose of fighting against terrorism or serious crime (by analogy C-203/15, Tele2). There must be appropriate safeguards protecting people concerned from possible abusive uses of biometric surveillance. Moreover, there must be effective legal remedies available to people regarding the use of biometric surveillance. Authorisations for targeted biometric surveillance must be subject to effective review by a court or an independence administrative body who has the power to issue legally binding decisions to verify that a situation justifying the recourse to the measure exists and the conditions and safeguards are observed (C‑511/18, C‑512/18 and C‑520/18, La Quadrature du Net and others, para 179).

A fundamental rights assessment of conducting targeted biometric surveillance in public spaces must be carried out in each stage of the data lifecycle, including when the data is processed in near real-time before collection. Especially where the personal data captured in the public sphere in real-time involves the use of data that the individual may not foresee (Uzun v Germany, para 45), that real-time automated processing would trigger an Article 8 protection. Similarly in the context of the right to personal data protection, the CJEU found in La Quadrature du Net and others (C‑511/18, C‑512/18 and C‑520/18) that the automated analysis of personal data amounted to an interference with the right to protection of personal data as set out in Article 8 of the Charter, even though it did not initially involve the collection of the data (para 170). Based on the case law of both courts, the automated analysis of biometric data in and of itself amounts to an interference with the right to privacy and personal data protection and must meet the fundamental rights requirements to be lawful. Accordingly, it must be subject to review by a court or an independent administrative body, and the pre-established tools or models used in the automated analysis must meet certain qualities (e.g., they must non-discriminatory, specific, and reliable; any positive result must be subject to manual and individual re-examination) (C‑511/18, C‑512/18 and C‑520/18, La Quadrature du Net and others, paras 180-182).

“Biometric data” in GDPR & LED

The General Data Protection Regulations (GDPR) provides the rules relating to the processing of personal data for all purposes except where the processing is carried out for the prevention, investigation, detection, or prosecution of criminal offences including the safeguarding against and the prevention of threats to public safety pursuant to its Article 2(2)(d). The Law Enforcement Directive (LED) complements the GDPR in this area as it applies specifically to the processing of personal data by competent authorities for the prevention, investigation, detection, or prosecution of criminal offences including the safeguarding against and the prevention of threats to public safety pursuant to its Article 1. 

Both legislations provide a specific framework for the processing of special categories of data (formerly known as “sensitive data”) – including biometric data, which is defined as “personal data” resulting from specific technical processing relating to the physical, physiological or behavioural characteristics of a natural person, which allow or confirm the unique identification of that natural person, such as facial images or “dactyloscopic data” under Article 4(14) of the GDPR and Article 3(13) of the LED. The definition thus recognises expanding categories of biometric data that can capture and measure human characteristics as it covers physical and physiological as well as behavioural biometric data. Notably, biometric data is granted a higher protection than non-sensitive personal data irrespective of the fact that they may not reveal sensitive information such as racial or ethnic origin, health, or sexual orientation.

Distinguishing personal data and biometric data

There are two elements that need to be sought for the personal data to constitute biometric data and for their processing to be subject to the specific limitations imposed by the GDPR and the LED.

- “Specific technical process”. Neither legislation defines the concept “specific technical process” but it should be understood as a special type of processing that captures the digital representation of biometric characteristics (e.g., facial images, fingerprints, voice) (Kindt 2013, 43; Jasserand 2016, 303). On this point, the European Data Protection Board (EDPB, 2019) notes that biometric data are the result of measurement of physical, physiological, or behavioural characteristics of individuals and thus the result of this special type of processing is captured by the concept of biometric data. For example, the image of the person captured by video surveillance is personal data, but it would be classified as biometric data once it is subjected to a specific type of processing to deduce the characteristics of that person (Recital 51, GDPR).

- “Unique identification of an individual”. Compared to the definition of personal data, it is unclear whether the element of identification for the purpose of defining biometric data requires a higher threshold (Jasserand 2013, 305-306). Both legislations define personal data broadly, as “any information relating to an identified or identifiable individual”. It has been confirmed both by the former Article 29 Data Protection Working Party (2007) and the CJEU (C-582/14, Breyer) that the personal data is broadly defined to capture the concept of “identifiability” whereby a person could be identifiable combined with other information available (including the information retained by someone other than the data controller) even if the person is not prima facie identified (paras 39-49).

The element of identification in the definition of biometric data on the other hand may suggest that said data must relate to an identified individual. The fact that the person could be identifiable through possible means would not be sufficient for the personal data to be classified as biometric data (Jasserand 2013, 306). The EDPB (2019) supports this view as it notes that if video surveillance system is set to detect the physical characteristics of individuals to classify them as opposed to uniquely identify them, this processing would not be subject to the framework reserved for the processing of sensitive data. Nevertheless, the data captured might still amount to personal data irrespective of the fact that they are not subject to any special type of processing.

Sensitive (biometric) data processing conditions

 Both the GDPR and the LED impose a special framework for the processing of sensitive data including biometric data as opposed to non-sensitive personal data. In essence, they impose limitations on the processing of sensitive data by setting out exceptional conditions for which the data may be processed. The following section considers the conditions under the LED on the processing of biometric data in order to set out the regulatory obligations relevant to implementing biometric surveillance in public spaces for law enforcement purposes. The conditions for biometric data processing under the GDPR are excluded from the scope of this report because it does not apply to processing activities for law enforcement purposes.

Active promotion

A certain number of actors, both at the national and at the local level are pushing for the development and the extension of biometric remote identification. At the local level, the new technological developments meet a growing apetite for smart city initiatives and the ambitions of mayors that strive for developing digital platforms and employ technology-oriented solutions for governance and law enforcement. The intention of the mayor of Nice, Christian Etrosi, to make Nice a “laboratory” of crime prevention, despite repeated concerns of the French DPA, is a case in point (for a detailed analysis, see chapter 8 in this report, see also Barelli 2018). Law enforcement agencies across Europe also continue to press ahead with efforts to build digital and automated infrastructures that benefits tech companies who push their face recognition technologies with the concept of smart city and innovation tech (ex. Huawei, NEC, etc.).

At the national level, Biometric systems for the purposes of authentication are increasingly deployed for forensic applications among law-enforcement agencies in the European Union. As we elaborate in Chapter 3, 11 out of 27 member states of the European Union are already using facial recognition against biometric databases for forensic purposes and 7 additional countries are expected to acquire such capabilities in the near future. The map of the European deployments of Biometric Identification Technologies (see Chapter 3) bear witness to a broad range of algorithmic processing of security images in a spectrum that goes from individual, localised authentication systems to generalised law enforcement uses of authentication, to Biometric Mass Surveillance.

Several states that have not yet adopted such technologies seem inclined to follow the trend, and push further. Former Belgian Minister of Interior Pieter De Crem for example, recently declared he was in favour of the use of facial recognition both for judicial inquiries but also for live facial recognition, a much rarer instance.

Support with safeguards

A second category of actors has indeed adopted the point of view that the RBI technologies should be supported, to the condition that their development should be monitored because of the risks they potentially pose. We find in this category the EU Commission, the EU Council, some EU Political parties, as well as the Fundamental Rights Agency (FRA), national DPAs such as the CNIL, the CoE (Council of Europe), and a certain number of courts.

Developments in the field of AI for governance, security and law enforcement are widely encouraged and financially supported by EU institutions. In their communication Shaping Europe’s Digital Futures accompanying the White Paper on AI, the European Commission set out its guidelines and strategies to create a “Europe fit for the digital age” (European Commission 2020a). In support of a “fair and competitive economy” the Commission proposes a European Data Strategy (EDS) to make Europe a global leader in the data-agile economy. The EDS further aims to ensure Europe’s technological sovereignty in a globalised world and “unlock the enormous potential of new technologies like AI” (Newsroom 2020). Therefore, the Commission proposes, among others “building and deploying cutting-edge joint digital capacities in the areas of AI, cyber, super and quantum computing, quantum communication and blockchain;” as well as “[r]einforcing EU governments interoperability strategy to ensure coordination and common standards for secure and borderless public sector data flows and services.” (European Commission 2020a, 4)

The financial support for these initiatives is planned to be channelled from the Digital Europe programme (DEP), the Connecting Europe Facility 2 and Horizon Europe. Through the Horizon Europe for instance, the Commission plans to invest €15 billion in the ‘Digital, Industry and Space’ cluster, with AI as a key activity to be supported. The DEP would benefit from almost €2.5 billion in deploying data platforms and AI applications while also supporting national authorities in making their high value data sets interoperable (Newsroom 2020).

In the European Parliament, the EPPEuropean People's Party most aligns with this approach. “We want to regulate facial recognition technologies, not ban them. We need clear rules where they can be used and where they must not be used”, has for example declared Emil Radev MEP, EPP Group Member of the Legal Affairs Committee. As he puts it “Without a doubt, we want to prevent mass surveillance and abuse. But this cannot mean banning facial recognition all together. There are harmless and useful applications for facial recognition, which increase personal security" (European People’s Party, 2021)

The FRA’s 2019 report on facial recognition technologies (FRA 2019), which builds on several previous reports concerning biometrics, IT systems and fundamental rights (FRA 2018); big data and decision making (FRA 2018); data quality and artificial intelligence (FRA 2019); calls for a moderate approach. The FRA advocates for a comprehensive understanding of how exactly facial recognition technologies work and what their impact on fundamental human rights are. Fundamental rights implications of using FRT, they argue, vary considerably depending on the purpose, scope and context. They highlight a number of issues based on the EU fundamental rights framework as well as the EU data protection legislation. For example, according to Article 9 of the GDPR, processing of biometric data is allowed based on the data subject’s explicit consent, which requires a higher threshold of precision and definitiveness including for processing purposes. In terms of using biometric surveillance in public spaces, explicit consent would not provide a lawful ground for the relevant data processing because– as observed by the CJEU in its Schwarz decision, the data subject who is entering the premises would not have any choice of opting out of data processing. If the processing of biometric data is based on substantial public interest, which is another lawful data processing ground under Article 9 of the GDPR, it must be “proportionate to the aim pursued, respect the essence of the right to data protection and provide for suitable and specific measures to safeguard the fundamental rights and interest of the data subjects” ((Article 9(2)(g), GDPR). Finally, when emphasising that the processing must be based on a lawful ground as recognised under the EU data protection legislation, the FRA was particularly vocal about the “function creep”, in regard to use of facial recognition systems and emphasised that the purpose of information collection must be strictly determined in light of the gravity of the intrusion upon people’s fundamental rights (25). 

Therefore, the FRA places the right to privacy and protection of personal and sensitive data at the core of their problem definition, emphasising the potential dangers of FRTs undermining the freedom of expression, association and assembly. The FRA report also makes a case for the rights of special groups such as children, the elderly and people with disabilities, and addresses the issue of how the use of FRTs can contribute to further criminalise and stigmatise already discriminated groups of people (e.g., certain ethnic or racial minorities). In light of these considerations they advocate for a clear and “sufficiently detailed” legal framework, close monitoring and a thorough and continuous impact assessment of each deployment.

The French DPA, the CNIL, takes a similar position in the report “Facial Recognition. For a debate living up to the challenges” (CNIL 2019b). The CNIL report argues that the contactless and ubiquitous nature of the different FRTs can create an unprecedented potential for surveillance which, in the long run, could potentially undermine societal choices.  They also emphasise that biometric data is sensitive data therefore its collection is never completely harmless: “Even legitimate and well-defined use can, in the event of a cyber-attack or a simple error, have particularly serious consequences. In this context, the question of securing biometric data is crucial and must be an overriding priority in the design of any project of this kind” (CNIL 2019b, 6). In their recommendations, while calling for special vigilance, they acknowledge the legitimacy and proportionality of some uses. The CNIL pointed out that GDPR-endangering applications are often presented as “pilot projects”, and thus requested the drawing of “some red lines even before any experimental use”. They call instead for “a genuinely experimental approach” that test and perfect technical solutions that respect the legal framework (CNIL 2019b, 10).

The CoE’s Practical Guide on the Use of Personal Data in the Police Sector (Council of Europe 2018), supplementing Convention 108+, puts great emphasis on implementing specific safeguards where an automated biometric system is introduced and considers that due to the high risk that such system poses to individuals’ rights, data protection authorities should be consulted in its implementation (10). Also, as mentioned below, the Council of Europe’s Guidelines on Facial Recognition (Council of Europe 2021), while considering a moratorium on the live facial recognition technology, sets out certain requirements to be met when implementing (possibly forensic) facial recognition technology.

Moratorium

On 20 January 2021,  the European Parliament (2021) passed a resolution on artificial intelligence in which they invite the Commission “to assess the consequences of a moratorium on the use of facial recognition systems, and, depending on the results of this assessment, to consider a moratorium on the use of these systems in public spaces by public authorities and in premises meant for education and healthcare, as well as on the use of facial recognition systems by law enforcement authorities in semi-public spaces such as airports, until the technical standards can be considered fully fundamental rights-compliant, the results derived are non-biased and non-discriminatory, and there are strict safeguards against misuse that ensure the necessity and proportionality of using such technologies;” (European Parliament 2021).

Another authority calling for a moratorium on automated recognition technologies in public spaces is the European Data Protection Supervisor (EDPS), the independent supervisory authority with responsibility for monitoring the processing of personal data by the EU institutions and bodies. According to their 2020-2024 Strategy (EDPS 2020) “Shaping a Safer Digital Future” released on 30 June 2020, the EDPS stresses that they are committed to supporting the idea of a moratorium on “the deployment, in the EU, of automated recognition in public spaces of human features, not only of faces but also of gait, fingerprints, DNA, voice, keystrokes and other biometric or behavioural signals, so that an informed and democratic debate can take place” (EDPS 2020).

The EDPS was also among the first to react to the draft Artificial Intelligence Act of the European Commission. While they welcomed the EU’s leadership aiming to ensure that AI solutions are shaped according to the EU’s values and legal principles, nonetheless they expressed their regret to see that their call for a moratorium on the use of remote biometric identification systems - including facial recognition - in publicly accessible spaces had not been addressed by the Commission. A stricter approach is necessary, they argue because “remote biometric identification, where AI may contribute to unprecedented developments, presents extremely high risks of deep and non-democratic intrusion into individuals’ private lives” (EDPS 2021). As mentioned below, shortly after their first reaction, the EDPS called for a general ban on the use of remote biometric systems with the European Data Protection Board (EDPB) (2021a).

A call for a moratorium, particularly, on facial recognition systems can be found in the Council of Europe documents. The Guidelines on Facial Recognition (Council of Europe, 2021) that is one of the instruments supplementing Convention 108+ call for a moratorium for the live facial recognition technologies (5) and lay out certain conditions for the use of facial recognition technologies by law enforcement authorities (6). For example, the Guidelines call for clear parameters and criteria when creating databases such as watchlists in light of a specific, legitimate, and explicit law enforcement purposes (ibid.)

Outright Ban

Finally, a certain number of EU Political Parties, EU and national NGOs have argued that there is no acceptable deployment of RBI, because the danger of Biometric Mass Surveillance is too high. Such actors include organisations such as EDRi, La Quadrature du Net, Algorithm Watch or the French Défenseur des Droits³⁴.

In the European Parliament, the European Greens have most vocally promoted the position of the ban, and have gathered support across party lines. In a letter to the European Commission dated 15 April 2021, 40 MEPs from the European Greens, the Party of the European Left, the Party of European Socialists, Renew Europe, a few non-attached MEPs and one member of the far-right party Identity and Democracy expressed their concerns about the leaked EU commission’ proposal for the AI Regulation a few days earlier. As they argued

The EU Commission Proposal

 In April 2021, the European Commission (2021b) published its proposal on the Regulation for the Artificial Intelligence Act with the aim of setting out the harmonised regulatory rules for Member States on AI- based systems. It responded in part to the many challenges posed by the rapid technological development of AI as well as the pressure from watchdogs, regulatory bodies and civil society. If adopted in its current form, the proposed EU Artificial Intelligence Act will have important implications for the use of biometric identification systems for law enforcement purposes.

On the whole, the proposed EU Artificial Intelligence Act lays out those rules based on three categories of possible risks that the use of AI may create: (i) an unacceptable risk according to which the use of AI is prohibited (Article 5); (ii) a high-risk AI system, whose use is subject to certain conditions including an ex-ante conformity assessment (Article 6); and (iii) low or minimal risk, whose use is permitted without restrictions.

Notably for the purpose of this report, the proposed EU Artificial Intelligence Act covers “remote biometric identification systems” defined as “an AI system for the purpose of identifying natural persons at a distance through the comparison of a person’s biometric data with the biometric data contained in a reference database, and without prior knowledge of the user of the AI system whether the person will be present and can be identified” (Article 3(36)). In this way, the proposed EU Artificial Intelligence Act anticipates covering (AI-based) biometric video surveillance systems. In so doing, it differentiates between the use of “real-time” and “post” remote biometric identification systems in public spaces for law enforcement purposes.

On initial observation, the proposal prohibits the use of “real-time” (live) remote biometric identification systems in public spaces for law enforcement purposes because it classifies them as systems that create an unacceptable risk.

However, Article 5 of the proposed EU Artificial Intelligence Act reads more as a heavy regulation rather than a prohibition. This is because the real-time remote biometric identification systems is prohibited unless it is “strictly necessary” for: (i) targeted search for specific potential victims of crime, including missing people; (ii) the prevention of a specific, substantial and imminent threat to the life or physical safety of natural persons or of a terrorist attack; or (iii) in relation to a criminal offence for which a European Arrest Warrant can be issued provided that it is punishable by a custodial sentence or detention order of minimum three years. In determining the use of real-time remote biometric identification systems for one of those purposes, Member States should be subject to appropriate limits in time, space, and target person (Article 5(2)). A court or an independent administrative body should authorise the use of this type of biometric identification systems, except in duly justified emergency situations (Article 5(3)). Member States may allow for full or partial use of real-time biometric identification systems in public spaces for law enforcement purposes based on the requirements laid out in Article 5 of the Proposed Regulation (Article 5(4)).

The use of “post” (forensic) remote identification systems for law enforcement purposes, on the other hand, is considered a high-risk AI system whose developers have the obligation to ensure that the system meets the condition set out in the proposed EU Artificial Intelligence Act (Chapter 2 and Annex III). As opposed to other high-risk AI systems whose providers have to conduct internal control checks, the post remote identification systems would be subject to third party conformity assessment.

The above provisions concerning the use of remote biometric identification systems have important implications on the protection of personal data and privacy as those systems involve processing of personal data. For this reason, they should be read alongside the rules and obligations set out in the GDPR and the LED. When conducted for law enforcement purposes by competent authorities, the remote biometric systems must follow the relevant data protection legislation as well as the ECHR and the Charter requirements since they would involve processing of sensitive personal data.

Reactions to the proposal

While data protection authorities and civil society generally welcomed the Commission’s initiative praising its horizontal approach and the broad scope of its application, several organisations expressed their concerns that the regulations put forward are often far too lenient and do not do enough to safeguard fundamental rights. The EDPS (2021) for instance voiced its concern that the Commission did not address their call for a moratorium on the use of biometric identification, and specifically on facial recognition systems. Notably, in their later opinion on the proposed EU Artificial Intelligence Act with the EDPB (EDPS and EDPB 2021), they called for a general ban on the use of biometric identification (as opposed to their earlier calls on its moratorium). Both institutions were particularly vocal about the high regulation as opposed to the prohibition of the use of real-time remote biometric identification and they observed that the conditions for which the system could be implemented were extensive and would render the so-called prohibition meaningless (11).

They were also very critical of the distinction between real-time and ex-post (forensic) use of biometric identification systems, noting that the latter is as intrusive as the former because of its chilling effect on the freedom of expression, of assembly, of association and the freedom of movement (12). Furthermore, they highlighted the inherently intrusive nature of all types of remote biometric identification systems as they would involve indiscriminate and disproportionate amount of personal data of data subjects in public spaces to identify a few individuals (ibid). This would consequently impact people’s reasonable expectation of anonymity in public spaces (ibid). For these reasons, the EDPS and the EDPB called for a general ban on ‘any use of AI for an automated recognition of human features in publicly accessible spaces’ (ibid).

The EDRi called the proposed EU Artificial Intelligence Act “a glimmer of hope for civil society who have called for red lines on impermissible uses of AI, like facial recognition in public spaces and predictive policing”, however, defined the prohibitions proposed as “relatively feeble”. They criticised in particular the many exemptions in which law enforcement agencies are still allowed to use “real-time remote biometric identification systems” (like facial recognition) in public spaces. These exceptions are targeted searches for specific potential victims of crime, to prevent a threat to life or terrorist attack, or to detect, localise, identify or prosecute a perpetrator or suspect of certain serious crimes (European Commission 2021b, 22). EDRi also points out that the Act “risks giving a green light to governments or public authorities to deploy discriminatory surveillance systems. Its rules for “high risk” AI – like predictive policing, AI in asylum procedures and worker surveillance – fall mainly on the developers themselves, not the public institutions actually deploying them – a cause for concern.” (EDRi 2021)

In June 2021 EDPB-EDPS also joined civil society in their call for a ban of automated facial recognition technologies (EDPB 2021a). In their joint opinion on the draft AI regulation (EDPB 2021a), they voice their concern of the exclusion of international law enforcement cooperation from the scope of the AI Proposal. While the EDPB and the EDPS welcome the risk-based approach underpinning the Proposal, they consider the concept of “risk to fundamental rights” as one which should be aligned with the EU data protection framework and the societal risks for groups of individuals should also be assessed and mitigated. Therefore, they call for “a general ban on any use of AI for automated recognition of human features in publicly accessible spaces, such as recognition of faces, gait, fingerprints, DNA, voice, keystrokes and other biometric or behavioural signals, in any context.” (EDPB 2021a))