Hubbry Logo
Computer vision dazzleComputer vision dazzleMain
Open search
Computer vision dazzle
Community hub
Computer vision dazzle
logo
7 pages, 0 posts
0 subscribers
Be the first to start a discussion here.
Be the first to start a discussion here.
Computer vision dazzle
Computer vision dazzle
Computer vision dazzle
View on Wikipedia
from Wikipedia

Computer vision dazzle, also known as CV dazzle, dazzle makeup, or anti-surveillance makeup, is a type of camouflage used to hamper facial recognition software, inspired by dazzle camouflage used by vehicles such as ships and planes.[1]

Methods

[edit]

CV dazzle combines stylized makeup, asymmetric hair, and sometimes infrared lights built in to glasses or clothing to break up detectable facial patterns recognized by computer vision algorithms in much the same way that warships contrasted color and used sloping lines and curves to distort the structure of a vessel.[2][3]

It has been shown to be somewhat successful at defeating face detection software in common use, including that employed by Facebook.[4][5] CV dazzle attempts to block detection by facial recognition technologies such as DeepFace "by creating an 'anti-face'".[6] It uses occlusion, covering certain facial features; transformation, altering the shape or colour of parts of the face; and a combination of the two.[7] Prominent artists employing this technique include Adam Harvey[8][9][10] and Jillian Mayer.[11]

Use in protests

[edit]

Computer vision dazzle makeup has been used by protestors in several different protest movements.[12] Its use as a protesting aid has often been found ineffective. It may be effective to thwart computer technology, but draws human attention, is easy for human monitors to spot on security cameras, and makes it hard for protestors to blend in within a crowd. Advances in facial recognition technology make dazzle makeup increasingly ineffective.[13]

See also

[edit]

References

[edit]
Revisions and contributorsEdit on WikipediaRead on Wikipedia

Computer vision dazzle

View on Grokipedia
from Grokipedia
Computer vision dazzle, commonly abbreviated as CV Dazzle, is an adversarial camouflage technique that employs bold, asymmetric patterns in makeup, hairstyling, and accessories to disrupt facial detection and recognition algorithms in systems by confounding feature extraction and contrast-based processing. Originating as an artistic and technical project by Adam Harvey in 2010 during his master's thesis at New York University's Interactive Telecommunications Program, CV Dazzle adapts principles from naval —which used disruptive geometries to mislead human estimation of speed and direction—to instead target vulnerabilities, prioritizing interference over concealment from human sight. The method operates by inverting luminance contrasts around facial landmarks (such as eyes and nose bridges), introducing artificial asymmetries, and fragmenting holistic face continuity, which lowers detection probabilities in algorithms like the Viola-Jones haarcascade classifier prevalent at the time of its inception. Early evaluations confirmed its capacity to evade such legacy systems, with Harvey's designs blocking detection in controlled tests from 2010 to 2016, and later iterations extending to convolutional neural networks through refined patterns exhibited in projects like Designs for a Different Future. However, peer-reviewed analyses have highlighted limitations against robust modern models, noting that CV Dazzle and similar physical perturbations often fail under real-world variations in lighting, angles, or algorithm retraining, rendering it more demonstrative of early adversarial fragility than a reliable . Despite these constraints, the project has notably advanced discourse on privacy-preserving interventions, inspiring extensions into textiles and wearable adversarial patterns while underscoring the cat-and-mouse dynamics between tactics and evolving technologies.

History

Origins in academic research

CV Dazzle originated as an academic project developed by artist and researcher Harvey during his master's at New York University's Interactive Program in 2010. The project focused on creating camouflage techniques to evade computer vision systems, particularly face detection algorithms, through asymmetric makeup patterns inspired by principles of disruption rather than concealment. Harvey's work was supervised within the program's emphasis on interactive media and technology, marking an early intersection of , concerns, and computational adversarial methods. The initiative emerged amid growing deployment of facial recognition technologies in consumer applications, including Facebook's introduction of automated photo-tagging features powered by facial recognition on December 15, 2010. Harvey aimed to prototype personal-level countermeasures, exploring how individuals could disrupt algorithmic without relying on institutional or legal interventions. This motivation reflected broader anxieties over proliferation in , where systems like Facebook's Tag Suggestions trained on vast user-generated datasets to identify and suggest tags for faces in images. Early dissemination included Harvey's project website, which showcased initial prototypes and pattern designs tested against commercial face detection software available in . He presented the work at the 2010 Next HOPE conference, demonstrating how the patterns could confuse detection models through visual asymmetry and feature occlusion. By 2011, interviews and features in outlets like highlighted the project's prototypes, positioning it as a foundational experiment in adversarial evasion.

Connection to historical dazzle camouflage

In 1917, British artist and naval officer Norman Wilkinson developed , applying bold, high-contrast geometric patterns to Allied ships to confound German observers' judgments of a vessel's speed, heading, and distance. These designs, influenced by cubist art, rejected concealment in favor of perceptual disruption, creating false cues about orientation and motion through clashing angles and stripes that complicated range-finding and aiming calculations. Although dazzle patterns demonstrably distorted human estimates of moving objects' direction and in settings, no rigorous wartime data confirmed they reduced shipping losses against submarine attacks, with post-war analyses indicating limited practical impact amid evolving threats like improved torpedoes. CV Dazzle adapts this misdirection paradigm to digital , explicitly drawing from Wilkinson's WWI technique to generate facial patterns that evade detection by algorithms, such as those relying on Haar-like features for identification. Where historical dazzle exploited biological vision's sensitivity to edge contrasts and for human misperception, CV Dazzle targets machine learning's dependence on statistical , yielding erroneous classifications without rendering the face imperceptible—mirroring the original's aim to confound rather than obscure. This lineage underscores a core perceptual divergence: human sight integrates contextual Gestalt principles vulnerable to geometric interference, while algorithmic processing parses discrete keypoints amenable to targeted perturbation, enabling analogous disruption across sensory modalities.

Technical foundations

Facial recognition algorithms targeted

Early facial recognition systems prevalent before 2010, such as the Viola-Jones framework introduced in , depended heavily on Haar cascade classifiers for detecting features through simple rectangular intensity contrasts corresponding to landmarks like eyes, bridge, and mouth. These classifiers operated in a multi-stage cascade, with each stage using boosted weak learners to evaluate Haar-like features—differences in pixel sums across adjacent regions—that exploit the high-contrast edges and symmetric geometry typical of human faces, such as darker regions around eyes against brighter cheeks. The algorithm's efficiency stemmed from rapid rejection of non-face regions via increasingly stringent tests, but this reliance on consistent spatial relationships and tonal gradients made it prone to failure when key features were obscured or asymmetrically perturbed. Disruptions targeting these vulnerabilities included alterations that reversed expected contrasts or introduced extraneous edges, overloading the probabilistic scoring in early cascade stages and preventing progression to later verification. For example, empirical demonstrations showed that perturbations avoiding or mimicking Viola-Jones detection could evade face localization entirely, as the assumes standardized proportions and without robustness to deliberate in high-importance regions like the symmetric eye-nose configuration. This susceptibility arose from the algorithm's on aligned, frontal faces, where deviations in reduced detection rates by confusing feature discriminators designed for holistic edge patterns rather than adversarial variability. Complementing detection, pre-2010 recognition methods like eigenfaces, proposed by Turk and Pentland in 1991, utilized to decompose face images into a low-dimensional subspace spanned by eigenvectors ("eigenfaces") capturing variance in training sets of symmetric, normalized faces. Recognition involved projecting query images onto this space and measuring to known templates, assuming perturbations would minimally affect projections within the manifold of typical facial geometry. However, asymmetric modifications disrupted this by shifting projections outside the learned subspace, as the method's sensitivity to global intensity variations and lack of invariance to non-linear deformations allowed bold, localized contrasts to inflate reconstruction errors and misclassify identities.

Principles of disruption in CV Dazzle

CV Dazzle patterns disrupt facial detection algorithms by exploiting vulnerabilities in feature-based classifiers, such as the Viola–Jones method, which relies on Haar-like features to identify symmetric arrangements of facial edges and contrasts in grayscale images. These algorithms, commonly implemented in 2010-era libraries via Haar cascade classifiers, scan for upright frontal faces using multi-stage boosting with 20–25 sequential classifiers that evaluate integral image features for rapid rejection of non-face regions. Asymmetric, high-contrast lines and geometric motifs break the bilateral algorithms expect in human faces, such as aligned eyes and , thereby failing the symmetry thresholds in early detection stages and inducing false negatives. Bold patterns occlude or exaggerate landmarks like the nose bridge, brow line, and eye contours—key anchors for bounding box initialization—preventing accurate localization and cascade progression. High-contrast elements overload edge detectors by flooding the image with extraneous gradients, diluting the of genuine facial edges and causing misprioritization in feature scoring. Reversal of expected luminance patterns, such as darkening symmetric highlights or lightening shadowed regions, further confuses the classifiers' assumptions about typical face , exploiting the fixed training priors of Haar cascades derived from datasets like the Viola–Jones training set. Open-source tests against OpenCV's haarcascade_frontalface_default, alt, and alt2 profiles demonstrated these mechanisms through saliency maps and detection failure rates on modified images. Unlike adversarial perturbations, which apply subtle, model-specific noise to force misclassification while remaining imperceptible to humans, CV Dazzle uses overt, human-applied designs prioritizing aesthetic viability and broad applicability over mathematical universality, targeting detection pipelines upstream of recognition to render faces undetectable rather than misidentified.

Implementation methods

Makeup and pattern designs

CV Dazzle employs bold, geometric makeup patterns designed to interfere with facial recognition algorithms by disrupting the detection of key landmarks such as the eyes, nose, and mouth. These patterns typically feature asymmetrical chevrons, stripes, and polygons applied around the eyes, cheeks, and nose bridge, utilizing high-contrast black and white elements to reverse natural light and dark areas on the face. This approach targets the assumptions of holistic face models, like the Viola-Jones algorithm, by breaking and feature continuity, thereby reducing the probability of automated detection. Application guidelines emphasize strategic placement on the T-zone (, , and ) and facial periphery to maximize interference with landmark detection while preserving basic wearability for human observers. Adam Harvey's 2010 master's thesis at New York University's Interactive Program outlined these techniques, recommending and unusual tonal directions—such as light makeup on darker skin or vice versa—to evade computational expectations without relying on obstructive accessories. Variations include color inversions and polygonal disruptions, tested through photographic demonstrations that achieved evasion in still images against primitive haarcascade classifiers (frontal, alt, and alt2 profiles) for multiple design iterations.

Accessories and extensions beyond cosmetics

Extensions of computer vision dazzle principles to non-cosmetic items emerged in the mid-2010s, focusing on scalable applications for broader disruption of surveillance systems. In 2017, artist and researcher Adam Harvey introduced the Hyperface project, which incorporates printed patterns resembling numerous tiny facial features—such as eyes, noses, and mouths—into fabrics for clothing and accessories. These designs aim to overload facial detection algorithms by generating thousands of false positives across the wearer's body, thereby diluting the system's focus on the actual face and reducing overall detection confidence scores. Unlike face-specific makeup, textile patterns enable coverage of larger areas, potentially interfering with full-body object detection or multi-view recognition in dynamic environments. Practical implementations include hoodies, shirts, and face masks embedded with adversarial motifs, as commercialized through platforms like Adversarial Fashion. Such items extend dazzle effects to garments that can be worn daily, targeting not only algorithms but also or recognition by scattering disruptive elements over the torso and limbs. Patterned hats and frames with asymmetric, high-contrast motifs have also been prototyped to obscure landmarks from overhead or side angles, building on core disruption tactics without relying on transient applications like . These accessories prioritize passive, durable interference, though scalability remains challenged by the need for precise alignment with camera perspectives and the higher material costs compared to . Empirical evaluations of these extensions, primarily from demonstrations, indicate partial against early commercial detectors like those used in retail , where false face detections can drop true positive rates by introducing into feature extraction pipelines. However, testing has been limited to controlled scenarios with specific algorithms, revealing vulnerabilities to viewpoint variations and the absence of rigorous, peer-reviewed benchmarks against evolving systems. This positions fabric and wearable extensions as conceptual advancements rather than foolproof countermeasures, with real-world deployment hindered by aesthetic compromises and inconsistent performance across diverse lighting and resolutions.

Practical applications

Adoption in protests and counter-surveillance activism

In the 2019 Hong Kong pro-democracy protests, demonstrators adopted strategies drawing from CV Dazzle principles, such as asymmetric face patterns and bold markings, to confound facial recognition systems deployed by authorities amid widespread surveillance. These tactics emerged as part of broader efforts to maintain in crowds, motivated by concerns over government tracking of participants via closed-circuit cameras and mobile scanning. During the 2020 George Floyd protests across the , CV Dazzle-inspired makeup gained traction among activists, with hashtags and tutorials promoting angular, high-contrast designs to obscure facial landmarks from police facial recognition tools. Participants applied these patterns to express resistance against perceived surveillance overreach, particularly in contexts where integrated automated identification for crowd monitoring and arrests. Privacy advocates and counter-surveillance groups have promoted CV Dazzle in activist circles as a non-violent tool for preserving during public assemblies, framing it as essential for exercising free speech without fear of retroactive identification. Coverage in outlets like has spotlighted such applications, portraying them as innovative extensions of historical dazzle tactics adapted for digital-era activism. Supporters emphasize protections, arguing that dazzle enables lawful in surveilled environments without relying on or hoods that might invite immediate scrutiny. Opponents, including some security commentators, counter that the conspicuous nature of these designs broadcasts evasion intent to human enforcers, potentially escalating risks for all protesters by associating the movement with deliberate concealment.

Commercial and artistic uses

CV Dazzle has been integrated into artistic exhibitions as that interrogates and identity through disruptive aesthetics. In 2018, featured the project in a spotlight on anti-surveillance makeup, presenting it as a method to confound facial recognition via bold, Cubist-inspired patterns applied to key features like the eyes, nose, and mouth. Similarly, at the Frankfurter Kunstverein, Adam Harvey's installation incorporated CV Dazzle styling advice to demonstrate evasion of automated identification systems, framing the technique as wearable critique of machine-readable visibility. These displays emphasize aesthetic experimentation over utility, positioning dazzle patterns as a medium for exploring the intersection of , , and . Harvey's ongoing artistic endeavors extend this intent, such as the 2023 CV Dazzle Anon Salon, a pop-up service applying makeovers to participants in controlled settings to highlight human-machine perceptual discrepancies. Through such interventions, the project functions less as activism and more as , influencing discourse in galleries and design circles on countering algorithmic gaze without broader societal mobilization. Commercial applications remain niche and underdeveloped, with no major makeup kits or apps directly commercializing CV Dazzle patterns despite post-2010 interest in privacy-focused . Discussions in outlets have speculated on its potential for marketable anti-detection products, yet the technique's specificity to outdated algorithms and abstract designs has constrained viability, yielding cultural resonance over sales. This limited uptake underscores CV Dazzle's primary role in raising artistic awareness of vulnerabilities rather than driving consumer markets.

Effectiveness evaluation

Empirical tests against early systems

In Adam Harvey's 2010 master's at New York University's Interactive Telecommunications Program, CV Dazzle patterns were tested against the Viola-Jones via OpenCV's Haar cascade classifiers (frontalface_default, frontalface_alt, frontalface_alt2, and profileface). Certain designs, generated using a to disrupt facial feature and contrast (e.g., eyes, nose bridge), achieved complete evasion of detection—up to 100% success—in controlled 2D still-image experiments under uniform lighting for all tested profiles, including profiles where initial designs failed. Independent evaluations in subsequent years, including a study by Wilber and Shmatikov, confirmed partial disruption against early-to-mid-2010s detectors, such as Facebook's , by inducing feature point errors that reduced detection rates, though exact evasion varied and was inconsistent across datasets like COFW (where baseline detection exceeded 97%). These tests emphasized efficacy in low-resolution, near-frontal views but highlighted failures in varied poses or higher-robustness models. Success remained algorithm-specific, primarily validated against Viola-Jones implementations prevalent in 2010-era systems, with no generalization to diverse classifiers; controlled conditions precluded real-world variability, and peer-reviewed randomized controlled trials establishing consistent blinding in uncontrolled environments were absent.

Limitations against modern computer vision

Modern systems, particularly those leveraging frameworks like convolutional neural networks (CNNs) developed post-2012 and architectures, demonstrate substantial resilience to CV Dazzle's disruptive patterns. These models, trained on expansive datasets augmented with variations in lighting, angles, and stylistic perturbations, extract hierarchical and invariant facial features that render asymmetric, high-contrast makeup designs largely ineffective for evasion. Unlike early hand-crafted detectors such as Viola-Jones cascades, which CV Dazzle targeted by exploiting specific feature asymmetries, contemporary systems employ and regularization techniques that mitigate the impact of such gross alterations. Empirical studies evaluating physical adversarial examples, including bold camouflage akin to CV Dazzle, report low success rates against robust facial recognition pipelines due to poor transferability from digital perturbations to real-world conditions. For example, designs from the project's inception around fail to confound current algorithms, as advancements in feature invariance and multi-scale processing bypass the intended fragmentation of . Quantitative assessments in systematizations of anti-facial recognition knowledge highlight that visible, non-subtle modifications achieve evasion in fewer than typical thresholds for practical utility, often inverting the goal by enhancing detectability under ensemble defenses. This obsolescence stems from causal factors in model training, where simulates dazzle-like distortions, training detectors to generalize beyond pattern-specific exploits. Recent analyses, including those against commercial deployments, affirm that CV Dazzle "likely doesn't work" reliably in post-2020 environments, prioritizing subtle, algorithm-tailored perturbations over overt ones for any residual efficacy.

Criticisms and debates

Practical drawbacks and human detectability

The bold, high-contrast patterns characteristic of CV Dazzle, including asymmetric facial markings and unconventional hairstyles, render wearers highly conspicuous to human observers, often amplifying rather than concealing their presence in public settings. This visibility stems from the deliberate disruption of and edges, which prioritizes algorithmic interference over naturalistic blending, thereby inviting scrutiny from bystanders or manual operators who may interpret the appearance as suspicious. For instance, empirical demonstrations in urban environments have shown that such designs fail to evade casual human gaze, as the aesthetics—evoking theatrical or artistic exaggeration—prompt social interactions or stigma, with wearers reported as standing out amid crowds. Wearability constraints further exacerbate operational flaws, as CV Dazzle relies predominantly on temporary that degrade with , environmental moisture, or , necessitating repeated application and rendering it impractical for extended or dynamic use. Prototypes often incorporate gender-specific elements like heavy or configurations suited to feminine features, limiting applicability and adoption among diverse demographics, including men, and introducing compliance challenges in non-static scenarios such as movement or varying lighting. Tests in real-world contexts, including settings, reveal inconsistent adherence, with patterns smudging or becoming illegible after minimal exposure to , thereby nullifying protective intent without human detectability mitigation. While advocates, including originator Adam Harvey, emphasize the technique's symbolic role in raising awareness about , observable behaviors in mixed human-AI monitoring environments demonstrate that heightened human perceptibility does not diminish overall identifiability, as manual verification can bypass automated failures through contextual cues like behavioral anomalies triggered by the attire. No verified data supports claims of reduced holistic surveillance efficacy from these designs, with critiques noting that the attention-drawing nature often prompts alternative tracking methods, such as following conspicuous individuals via non-facial means.

Broader societal and security implications

Facial recognition systems have contributed to measurable improvements in public safety by aiding in identifying suspects and reducing . A 2024 analysis of 268 U.S. cities found that police deployment of facial recognition applications correlated with reductions in felony violence and rates, without corresponding increases in arrests for non-violent offenses. The New York Police Department, for instance, integrates the with human verification to solve crimes more efficiently, emphasizing its role in enhancing investigative capabilities rather than replacing traditional methods. Leading systems achieve accuracy exceeding 99% in controlled benchmarks, supporting rapid suspect identification and prevention of threats. CV Dazzle and similar methods introduce risks by facilitating evasion of these systems, potentially shielding criminals from detection. Studies on adversarial accessories demonstrate that targeted perturbations can evade state-of-the-art facial recognition in up to 95% of tested cases, enabling "stealthy attacks" that preserve human imperceptibility while undermining algorithmic reliability. analyses highlight how such techniques could empower threats during high-stakes operations, as seen in broader biometric evasion research where modifications allow perpetrators to bypass watchlists without alerting human observers. While privacy advocates contend that facial recognition fosters pervasive surveillance and disproportionate impacts on marginalized groups, empirical data tempers claims of systemic oppression by revealing error rates that, in aggregate, favor low false positives for innocents in law enforcement contexts and underscore benefits like cold case resolutions. Bans on the technology in cities such as San Francisco (2019) and Boston (2020) reflect these concerns, yet subsequent crime surges have prompted some jurisdictions to reconsider restrictions, with no evidence linking prohibitions to curtailed government overreach. Prioritizing public safety through accurate, regulated deployment aligns with causal evidence of crime deterrence over unverified privacy gains from evasion tools like dazzle.

References

  1. https://adam.harvey.studio/cvdazzle
  2. https://walkerart.org/calendar/2020/designs-for-different-futures
  3. https://www.cs.cornell.edu/~shmat/shmat_wacv16.pdf
  4. https://arxiv.org/pdf/2412.13507
  5. https://dl.acm.org/doi/pdf/10.1145/3535192
  6. https://www.forbes.com/sites/mikeisaac/2010/12/15/facebook-to-add-facial-recognition-software-to-photo-tagging/
  7. https://www.cnn.com/2010/TECH/social.media/12/16/facebook.facial.recognition.mashable/index.html
  8. https://ahprojects.com/projects/cv-dazzle
  9. https://www.youtube.com/watch?v=I29V_HYdNH4
  10. https://daily.jstor.org/the-camouflage-that-dazzled/
  11. https://www.bbc.co.uk/teach/articles/zmkx8xs
  12. https://royalsocietypublishing.org/doi/10.1098/rsos.240624
  13. https://ieeexplore.ieee.org/document/990517
  14. https://www.cs.cmu.edu/~efros/courses/LBMV07/Papers/viola-cvpr-01.pdf
  15. https://users.ece.cmu.edu/~lbauer/papers/2016/ccs2016-face-recognition.pdf
  16. https://arxiv.org/pdf/1910.06261
  17. https://direct.mit.edu/jocn/article/3/1/71/3025/Eigenfaces-for-Recognition
  18. https://www.face-rec.org/algorithms/PCA/jcn.pdf
  19. https://docs.opencv.org/4.x/d2/d99/tutorial_js_face_detection.html
  20. https://www.youtube.com/watch?v=hPCTwxF0qf4
  21. https://www.wired.com/story/anti-surveillance-clothing-adam-harvey/
  22. https://adam.harvey.studio/hyperface/
  23. https://www.theguardian.com/technology/2017/jan/04/anti-surveillance-clothing-facial-recognition-hyperface
  24. https://adversarialfashion.com/
  25. https://boingboing.net/2017/01/04/hyperface-a-fabric-that-makes.html
  26. https://www.vice.com/en/article/countersurveillance-textiles-trick-computer-vision-software/
  27. https://www.cnn.com/style/article/protest-design-future-intl-hnk
  28. https://slate.com/technology/2019/08/facial-recognition-surveillance-fashion-hong-kong.html
  29. https://www.vogue.com/article/anti-surveillance-makeup-cv-dazzle-protest
  30. https://intothegloss.com/2020/06/cv-dazzle-protest-makeup/
  31. https://www.theguardian.com/world/2020/feb/01/privacy-campaigners-dazzle-camouflage-met-police-surveillance
  32. https://www.youtube.com/watch?v=p1yCw77ShkA
  33. https://www.vogue.com/article/computer-vision-dazzle-anti-surveillance-facial-recognition-technology-moma-ps1
  34. https://www.fkv.de/en/adam-harvey/
  35. https://www.theartblog.org/2020/01/adam-harvey-and-the-anti-face/
  36. https://www.diyphotography.net/cv-dazzle-makeup-system-kills-face-detection/
  37. https://olh.openlibhums.org/article/id/10181/
  38. https://people.cs.uchicago.edu/~ravenben/publications/pdf/sok-oakland23.pdf
  39. https://www.theregister.com/2025/01/15/make_up_thwart_facial_recognition/
  40. https://ironholds.org/madeup-resistance/
  41. https://www.freethink.com/culture/anti-facial-recognition
  42. https://resistance-journal.org/wp-content/uploads/2020/01/JRS-2-2019-Recognizing-Everyday-Activism.pdf
  43. https://greydynamics.com/razzle-dazzle-against-facial-recognition/
  44. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4796951
  45. https://www.nyc.gov/site/nypd/about/about-nypd/equipment-tech/facial-recognition.page
  46. https://www.securityindustry.org/2022/07/23/what-science-really-says-about-facial-recognition-accuracy-and-bias-concerns/
  47. https://dl.acm.org/doi/10.1145/2976749.2978392
  48. https://www.clearview.ai/post/the-myth-of-facial-recognition-bias
  49. https://www.npr.org/sections/live-updates-protests-for-racial-justice/2020/06/24/883107627/boston-lawmakers-vote-to-ban-use-of-facial-recognition-technology-by-the-city
  50. https://www.usccr.gov/files/2024-09/civil-rights-implications-of-frt_0.pdf
  51. https://www.sciencedirect.com/science/article/pii/S0264275124006863
Add your contribution
Related Hubs
User Avatar
No comments yet.