Scheda di revisione: Critique des algorithmes de classification de genre

📋 Plan du Cours

1. Indétermination de l'identité et de la vie des espèces
2. Interconnexion des caractéristiques physiques et des systèmes d'information
3. Analyse automatisée du visage et classification de genre
4. Évaluation des performances des classificateurs de genre commerciaux
5. Conclusion sur la précision des algorithmes de classification de genre

📖 1. Indétermination de l'identité et de la vie des espèces

🔑 Notions clés & Définitions

• Biometric borders : frontières qui utilisent des technologies biométriques pour gérer la mobilité, en intégrant des frontières physiques et symboliques, et en inscrivant le corps dans des encadrements multiples.
• Biometric border : limite qui repose sur la représentation biométrique du corps, servant à distinguer, contrôler et prévoir la comportement des individus.
• the biometric border : frontière biométrique qui étend la gestion de la mobilité à des domaines variés de la vie quotidienne, au-delà des frontières géographiques traditionnelles.
• such that : expression indiquant que la vie et l’identité sont des notions fluides, incertaines, et en constante transformation, rendant leur fixation difficile.
• Geography 25 (2006) 336e351 : référence à une publication qui analyse la complexité et l’indécidabilité de la vie des espèces face aux techniques biométriques.

📝 Points essentiels

• La vie des espèces est une donnée indécidable, non fixée définitivement, ce qui complique la gestion biométrique des frontières.
• L’identité n’est jamais totalement constituée, et l’identification biométrique ne peut réduire complètement l’écart entre identité réelle et représentation.
• Le concept de ‘bare life’ désigne une vie réduite à sa calculabilité, vulnérable à la violence et à la suspension des droits, notamment dans le contexte biométrique.
• La gestion biométrique des frontières ne se limite pas à la territorialité, mais s’étend à des frontières sociales, légales, raciales, et de genre, inscrites dans le corps et la vie quotidienne.

💡 À retenir

L’incertitude et la fluidité de l’identité et de la vie rendent la gouvernance biométrique intrinsèquement instable, révélant des tensions entre contrôle, violence et indécidabilité.

📖 2. Interconnexion des caractéristiques physiques et des systèmes d'information

🔑 Notions clés & Définitions

• Encoded risk profiles : profils de risque encodés utilisant des données numériques pour prédire et prévenir des actes futurs, notamment avant l’arrivée à la frontière.
• Preemptive fixing of identities : fixation anticipée des identités à partir de traces électroniques, permettant de prévoir et d’intervenir avant la rencontre physique.
• with information systems : intégration des caractéristiques physiques dans des systèmes numériques pour gérer la mobilité et l’identification.
• with information : utilisation de données électroniques et biométriques pour représenter, prévoir et contrôler l’identité et la mobilité des individus.

📝 Points essentiels

• Les profils de risque encodés sont utilisés pour prédire et prévenir des actes futurs en séparant, avant même la frontière physique, les mobilités légitimes (business, loisir) des mobilités illégitimes (terrorisme, immigration).
• Le système US VISIT fixe préventivement les identités à partir de traces électroniques (billets, visas, cartes de crédit), permettant une gestion anticipée des risques.
• La surveillance par données relie progressivement les caractéristiques physiques individuelles aux systèmes d’information, facilitant la gestion des mobilités et la catégorisation des individus.
• La technique d’encodage des risques permet de prévoir et d’intervenir sur la mobilité, en fixant des identités numériques à partir de traces électroniques, renforçant la capacité de prévention avant la rencontre physique.

💡 À retenir

Les caractéristiques physiques sont intégrées dans des systèmes d’information pour anticiper et contrôler les identités, permettant une gestion préventive des mobilités avant même leur manifestation physique.

📖 3. Analyse automatisée du visage et classification de genre

🔑 Notions clés & Définitions

• Face-based gender classification : classification automatique du genre à partir d’images faciales, influencée par la caractéristique du visage.
• gender classification : processus d’identification du genre d’un individu via des algorithmes, souvent à partir d’images ou vidéos.
• classification system : ensemble structuré d’algorithmes ou méthodes permettant de catégoriser des visages selon des critères comme le genre ou la peau.
• classification algorithms : programmes informatiques qui analysent des traits faciaux pour déterminer le genre ou d’autres caractéristiques.
• Skin Type : classification de la peau selon le système de Fitzpatrick, utilisée pour caractériser la pigmentation cutanée.

📝 Points essentiels

• La précision de la classification de genre varie selon le genre et le type de peau, avec des erreurs plus fréquentes chez les femmes à peau plus foncée.
• Les bases de données utilisées pour l’analyse faciale incluent des sous-ensembles diversifiés en pose, éclairage et résolution pour assurer la qualité des images.
• Les disparités dans les taux d’erreur indiquent que les algorithmes sont influencés par l’intersection du genre et du type de peau, avec des performances moindres pour les femmes à peau foncée.

💡 À retenir

Les performances des algorithmes de classification de genre sont fortement affectées par la diversité phénotypique, notamment la couleur de peau, ce qui soulève des enjeux d’équité et de fiabilité.

📖 4. Évaluation des performances des classificateurs de genre commerciaux

🔑 Notions clés & Définitions

• NIST Evaluation of Automated Gender Classification Algorithms : étude qui mesure la performance des algorithmes commerciaux de classification de genre en utilisant des datasets annotés selon des critères précis, notamment la couleur de peau.
• Commercial gender classifiers : systèmes de reconnaissance faciale vendus sous forme d’API, conçus pour déterminer le genre à partir d’images faciales, dont la performance varie selon la démographie.
• time of evaluation : période durant laquelle les algorithmes sont testés, influençant la représentativité et la performance mesurée.
• the time of evaluation : moment précis où la performance des classificateurs est mesurée, impactant la comparabilité des résultats.
• error rate : pourcentage d’erreurs de classification, indiquant la fiabilité d’un système, notamment plus élevé chez certains groupes démographiques.

📝 Points essentiels

• Les classificateurs commerciaux ont des performances inégales selon la couleur de peau, avec une précision supérieure sur les sujets à peau claire.
• Microsoft est identifié comme le meilleur classificateur dans l’étude, mais tous présentent des biais raciaux, notamment une erreur accrue chez les sujets à peau foncée.
• Les erreurs de classification sont plus fréquentes chez les sujets à peau foncée, en particulier pour les femmes, avec un taux pouvant atteindre 34,7 %, contre 0,8 % pour les sujets à peau claire.
• La performance est systématiquement meilleure sur les hommes que sur les femmes, et sur les sujets à peau claire que sur ceux à peau foncée.
• La disparité la plus marquée concerne la classification des femmes à peau foncée, nécessitant une attention urgente pour améliorer l’équité et la fiabilité.

💡 À retenir

Les classificateurs commerciaux de genre montrent des biais raciaux et de genre, avec des erreurs significatives pour les groupes à peau foncée, ce qui soulève des enjeux éthiques et pratiques pour leur usage dans des contextes sensibles.

📖 5. Conclusion sur la précision des algorithmes de classification de genre

🔑 Notions clés & Définitions

• Égalité des chances en apprentissage supervisé : situation où les algorithmes produisent des résultats équitables pour toutes les populations, sans biais démographique ou phenotypique.

• Reconnaissance faciale non régulée : utilisation de technologies de surveillance sans contrôle ou standards stricts, pouvant entraîner des discriminations ou erreurs systématiques.

• Classification de genre : tâche de l’analyse faciale visant à déterminer le sexe d’une personne à partir d’images, dont la fiabilité dépend de la qualité des données et de la diversité des populations.

• Reconnaissance faciale : processus d’identification ou de vérification d’individus via leur visage, susceptible de biais et d’erreurs selon les contextes et les populations.

• Analyse faciale : ensemble des techniques automatiques pour détecter, classifier ou interpréter des traits du visage, influencées par la qualité des données et la représentativité des échantillons.

📝 Points essentiels

• Les algorithmes de reconnaissance faciale et classification de genre présentent des biais affectant l’égalité des chances, notamment une moindre précision pour certains groupes démographiques comme les femmes ou les personnes de couleur.

• La performance des systèmes dépend fortement de la qualité et de la diversité des données d’entraînement, ce qui peut conduire à des erreurs accrues pour des populations sous-représentées.

• L’usage non régulé de ces technologies par la police soulève des enjeux majeurs de surveillance, de discrimination et de violation des libertés civiles, notamment par des taux plus élevés de fausses identifications chez certains groupes.

💡 À retenir

La fiabilité des algorithmes de classification de genre est conditionnée par la qualité des données et leur diversité, soulignant l’importance d’une régulation pour garantir une utilisation éthique et précise.

🧩 Compléments de couverture

1. Détail source à réviser : borders: Governing mobilities in the war on terror* Louise Amoore* Department of Geography, University of Durham, South Road, Durham DH1 3LE, UK Abstract This article proposes the concept of the biometric border in order (Source: "borders: Governing mobilities in the war on terror* Louise Amoore* Department of Geography, University of Durham, South Road, Durham DH1 3LE, UK Abstract This article proposes the concept of the biometric border in order to signal a dual-faced phenomenon in the contemporary war on terror: the turn to scientific technologies and managerial expertise")
2. Détail source à réviser : Ltd. All rights reserved. doi:10.1016/j.polgeo.2006.02.001 Political Geography 25 (2006) 336e351 www.elsevier.com/locate/polgeo and happiness through science and law. When they fail, this only justifies the need for more (Source: "Ltd. All rights reserved. doi:10.1016/j.polgeo.2006.02.001 Political Geography 25 (2006) 336e351 www.elsevier.com/locate/polgeo and happiness through science and law. When they fail, this only justifies the need for more of the same. (Dreyfus & Rabinow, 1983, p. 196). Had information coordination technology been properly in place before September 11,")
3. Détail source à réviser : management of the border cannot be understood simply as a matter of the geopolitical policing and disciplining of the movement of bodies across mapped space. Rather, it is more appropriately understood as a mat- ter of b (Source: "management of the border cannot be understood simply as a matter of the geopolitical policing and disciplining of the movement of bodies across mapped space. Rather, it is more appropriately understood as a mat- ter of biopolitics, as a mobile regulatory site through which people’s everyday lives can be made amenable to intervention and management. In this")
4. Détail source à réviser : In effect, the biometric border is the portable border par excellence, carried by mobile bodies at the very same time as it is deployed to divide bodies at international boundaries, airports, railway stations, on subways (Source: "In effect, the biometric border is the portable border par excellence, carried by mobile bodies at the very same time as it is deployed to divide bodies at international boundaries, airports, railway stations, on subways or city streets, in the office or the neighbourhood. The work of the biometric border is thus the work of redefining what Bigo")
5. Détail source à réviser : is not a datum, it is an undecidable’. Though the biometric border undeniably draws species life into the exercise of power, it is necessarily working with an unstable and unpredictable referent. Throughout this paper I (Source: "is not a datum, it is an undecidable’. Though the biometric border undeniably draws species life into the exercise of power, it is necessarily working with an unstable and unpredictable referent. Throughout this paper I will suggest ways in which the ambivalent, antagonistic and undecidable moments of the biometric border might be revealed. Though, as")
6. Détail source à réviser : becomes one of isolating the legitimate ‘inside’ transborder activities of the global economy, and securing them from the illegitimate ‘outside’ of those who would exploit the pos- sibilities of open borders. I have argu (Source: "becomes one of isolating the legitimate ‘inside’ transborder activities of the global economy, and securing them from the illegitimate ‘outside’ of those who would exploit the pos- sibilities of open borders. I have argued elsewhere, following Pat O’Malley and others, that the discursive deployment of risk, particularly by management consultants, is")
7. Détail source à réviser : entry and exit data; APIS, containing 339L. Amoore / Political Geography 25 (2006) 336e351 passenger manifest information; SEVIS, containing data on all foreign and exchange students in the United States; IBIS, a ‘lookou (Source: "entry and exit data; APIS, containing 339L. Amoore / Political Geography 25 (2006) 336e351 passenger manifest information; SEVIS, containing data on all foreign and exchange students in the United States; IBIS, a ‘lookout’ watch list interfaced with Interpol and national crime data; CLAIMS3, holding information on foreign nationals claiming")
8. Détail source à réviser : as ‘hits’ on the various databases, and how a ‘false hit’ that leads to detention or deportation can be challenged. 4 As one EPIC lawyer put the problem: ‘these technologies are assumed to provide a complete pic- ture of (Source: "as ‘hits’ on the various databases, and how a ‘false hit’ that leads to detention or deportation can be challenged. 4 As one EPIC lawyer put the problem: ‘these technologies are assumed to provide a complete pic- ture of who someone is, leaving people having to dispute their own identity’. 5 In these terms the US VISIT system far exceeds a technologized")
9. Détail source à réviser : it is perhaps in other spheres that we find instances of the politicisation of the techniques used at the biometric border. In common with others who have pointed to the capacity of the satirical and playful prac- tices (Source: "it is perhaps in other spheres that we find instances of the politicisation of the techniques used at the biometric border. In common with others who have pointed to the capacity of the satirical and playful prac- tices of the arts to disrupt our sense of the ‘normal run of things’ (Bleiker, 2000; Butler, 2004; De Goede, 2005), I am going to suggest")
10. Détail source à réviser : of electronic personal data in order to classify and govern the movement of people across borders has become a key feature of the contemporary war on terror. The US VISIT programme, though, extends the use of integrated (Source: "of electronic personal data in order to classify and govern the movement of people across borders has become a key feature of the contemporary war on terror. The US VISIT programme, though, extends the use of integrated personal data into biometrics, a move that signals what Levi and Wall (2004, p. 194) have termed a ‘new politics of surveil- lance’. To")
11. Détail source à réviser : can be safely secured. What van der Ploeg (2003, p. 58) has observed as a gradually extending intertwinement of in- dividual physical characteristics with information systems’ has served to deepen faith in data as a mean (Source: "can be safely secured. What van der Ploeg (2003, p. 58) has observed as a gradually extending intertwinement of in- dividual physical characteristics with information systems’ has served to deepen faith in data as a means of risk management and the body as a source of absolute identification. Biometric technologies are perhaps best understood as")
12. Détail source à réviser : off the shelf that wasn’t effective’ (New York Times, May 8, 2005). 7 Full text of speech is available at: www.useu.be/terrorism/EUResponse/Jan1305RidgeNetherlandsProgram.html. 342 L. Amoore / Political Geography 25 (200 (Source: "off the shelf that wasn’t effective’ (New York Times, May 8, 2005). 7 Full text of speech is available at: www.useu.be/terrorism/EUResponse/Jan1305RidgeNetherlandsProgram.html. 342 L. Amoore / Political Geography 25 (2006) 336e351 what is called ‘risk pooling’ in studies of the insurance industry (cf. Ewald, 1991; Heimer, 2002). By categorizing patterns")
13. Détail source à réviser : verifiers of the truth about a person e the ultimate guar- antors of identity. As such, they are increasingly being seen as the smart scientific solution to the problem of fighting the war on terror without impeding glob (Source: "verifiers of the truth about a person e the ultimate guar- antors of identity. As such, they are increasingly being seen as the smart scientific solution to the problem of fighting the war on terror without impeding globalization e the means of managing risk by embracing risk (Baker & Simon, 2002) or, in Dillon and Reid’s (2001) terms, of")
14. Détail source à réviser : The ever-present gap between identity and identification, or what is unrealizable in the discursive making of the subject, has been a preoccupation of social and cultural theory for some time. Despite radical differences (Source: "The ever-present gap between identity and identification, or what is unrealizable in the discursive making of the subject, has been a preoccupation of social and cultural theory for some time. Despite radical differences of ap- proach, there is some sense of valuing the ‘gap’ politically as a potential space for contestation and dissent. Since the")
15. Détail source à réviser : frequent fliers become accustomed to access and open borders, Bunting fleetingly confronts the viewer with something of ‘the everyday experience of illegal border crossers’. ‘Today’s borders are not so much about permiss (Source: "frequent fliers become accustomed to access and open borders, Bunting fleetingly confronts the viewer with something of ‘the everyday experience of illegal border crossers’. ‘Today’s borders are not so much about permission and refusal of entry as about user profiling’, reflects Bunting, ‘the ultimate aim being the filtering of presumably useful from")
16. Détail source à réviser : beyond vague concerns about an offshore company winning a US government contract.8 In effect, the expertise becomes the norm, as one immigration lawyer explained, ‘since 9/11 the public authorities have turned to the pri (Source: "beyond vague concerns about an offshore company winning a US government contract.8 In effect, the expertise becomes the norm, as one immigration lawyer explained, ‘since 9/11 the public authorities have turned to the private authorities to design the architecture of the systems, to make ‘‘efficient systems’’. so this is only ever treated as a technical")
17. Détail source à réviser : officials with a demo that included wireless tags that tracked immigrants’ whereabouts’ (‘‘Accenture Hits the Daily 8 Interviewed in the New York Times, Representative Richard Neal declared the US VISIT contact to be ‘ou (Source: "officials with a demo that included wireless tags that tracked immigrants’ whereabouts’ (‘‘Accenture Hits the Daily 8 Interviewed in the New York Times, Representative Richard Neal declared the US VISIT contact to be ‘outrageous’. ‘The Bush administration’, he argued, ‘has awarded the largest homeland security contract in history to a company that has")
18. Détail source à réviser : in the war on terror, ob- serving the behaviour of fellow passengers on a train, new neighbours in town, ‘and anyone who looks vaguely Arab in the dominant racial imaginary’. At the time of writing, in the wake of the Ju (Source: "in the war on terror, ob- serving the behaviour of fellow passengers on a train, new neighbours in town, ‘and anyone who looks vaguely Arab in the dominant racial imaginary’. At the time of writing, in the wake of the July London bombings, the Chief Constable of London transport police is calling for increased vigilance on the part of commuters and")
19. Détail source à réviser : I 346 L. Amoore / Political Geography 25 (2006) 336e351 have discussed in greater depth the problematic of resistance within the war on terror, suggest- ing that we find ourselves in ambivalent subject positions: both fr (Source: "I 346 L. Amoore / Political Geography 25 (2006) 336e351 have discussed in greater depth the problematic of resistance within the war on terror, suggest- ing that we find ourselves in ambivalent subject positions: both frequent flier and immigrant rights campaigner, for example; or both London city commuter and anti-war protester (Amoore, 2006). Though")
20. Détail source à réviser : 2002) puts it, and ‘learning to live’ with our fears. Indeed, as many writers have suggested, art, comedy and laughter have an important role to play in a politics that disrupts what we have come to see as necessary or n (Source: "2002) puts it, and ‘learning to live’ with our fears. Indeed, as many writers have suggested, art, comedy and laughter have an important role to play in a politics that disrupts what we have come to see as necessary or normal ways of living (Bleiker, 2000; Odysseos, 2001). In the face of a war on terror that appeals to our sense of normal ‘ways of life’")
21. Détail source à réviser : de- bate turned to questions of identity, status and profiling. In effect, once it became clear that de Menezes could not be represented as the terrorist embodiment of bare life, a struggle began to reposition his ‘other (Source: "de- bate turned to questions of identity, status and profiling. In effect, once it became clear that de Menezes could not be represented as the terrorist embodiment of bare life, a struggle began to reposition his ‘otherness’ as that of the illegal immigrant. The discovery that Jean Charles’ stu- dent visa may have expired two years previously led to")
22. Détail source à réviser : verifiable identity at the biometric border thus becomes a condition of being, in the sense of living within a particular society or way of life, if not indeed a condition of life itself. I have argued throughout this pa (Source: "verifiable identity at the biometric border thus becomes a condition of being, in the sense of living within a particular society or way of life, if not indeed a condition of life itself. I have argued throughout this paper that the biometric border implicates us all in the govern- ing of mobility and in the profiling of suspicious behaviour. It does so via")
23. Détail source à réviser : discipline at work. Economy and Society, 33(2), 174e196. Amoore, L. (2006). There is no great refusal: The ambivalent politics of resistance. In M. de Goede (Ed.), International political economy and poststructural polit (Source: "discipline at work. Economy and Society, 33(2), 174e196. Amoore, L. (2006). There is no great refusal: The ambivalent politics of resistance. In M. de Goede (Ed.), International political economy and poststructural politics. Basingstoke, UK: Palgrave. Amoore, L., & de Goede, M. (2005). Governance, risk and dataveillance in the war on terror. Crime, Law and")
24. Détail source à réviser : Excellence. (2004). We the people: Homeland security from the citizens’ perspective. Council for Excel- lence in Government in association with Accenture. Coutin, S. B. (2000). Legalizing moves: Salvadoran immigrants’ st (Source: "Excellence. (2004). We the people: Homeland security from the citizens’ perspective. Council for Excel- lence in Government in association with Accenture. Coutin, S. B. (2000). Legalizing moves: Salvadoran immigrants’ struggle for US residency. Ann Arbor: University of Michigan Press. Coutin, S. B., Maurer, B., & Yngvesson, B. (2002). In the mirror: The")
25. Détail source à réviser : P. Miller (Eds.), The Foucault effect: Studies in gov- ernmentality. Chicago: University of Chicago Press. Foucault, M. (1976). The will to knowledge. In: The history of sexuality, Vol. 1. London: Penguin. Foucault, M. ( (Source: "P. Miller (Eds.), The Foucault effect: Studies in gov- ernmentality. Chicago: University of Chicago Press. Foucault, M. (1976). The will to knowledge. In: The history of sexuality, Vol. 1. London: Penguin. Foucault, M. (1983). Afterword: The subject and power. In H. L. Dreyfus, & P. Rabinow (Eds.), Michel Foucault: Be- yond structuralism and hermeneutics.")
26. Détail source à réviser : Studies, 30(3), 709e732. O’Malley, P. (2000). Uncertain subjects: Risks, liberalism and contract. Economy and Society, 29(4), 460e484. Plan to improve relations with Muslim communities. (2005, August 1). The Guardian, 8. (Source: "Studies, 30(3), 709e732. O’Malley, P. (2000). Uncertain subjects: Risks, liberalism and contract. Economy and Society, 29(4), 460e484. Plan to improve relations with Muslim communities. (2005, August 1). The Guardian, 8. van der Ploeg, I. (2003). Biometrics and the body as information: Normative issues of the socio-technical coding of the body. In D. Lyon")
27. Détail source à réviser : present in automated facial analysis al- gorithms and datasets with respect to phe- notypic subgroups. Using the dermatolo- gist approved Fitzpatrick Skin Type clas- sification system, we characterize the gen- der and sk (Source: "present in automated facial analysis al- gorithms and datasets with respect to phe- notypic subgroups. Using the dermatolo- gist approved Fitzpatrick Skin Type clas- sification system, we characterize the gen- der and skin type distribution of two facial analysis benchmarks, IJB-A and Adience. We find that these datasets are overwhelm- ingly composed of")
28. Détail source à réviser : 2016; Caliskan et al., 2017). Bolukbasi et al. even showed that the popular word embedding space, Word2Vec, encodes soci- etal gender biases. The authors used Word2Vec to train an analogy generator that fills in miss- in (Source: "2016; Caliskan et al., 2017). Bolukbasi et al. even showed that the popular word embedding space, Word2Vec, encodes soci- etal gender biases. The authors used Word2Vec to train an analogy generator that fills in miss- ing words in analogies. The analogy man is to computer programmer as woman is to “X” was completed with “homemaker”, conforming to the")
29. Détail source à réviser : well as their use is necessary to protect citizens’ rights and keep vendors and law enforcement accountable to the public. We take a step in this direction by making two contributions. First, our work advances gender cla (Source: "well as their use is necessary to protect citizens’ rights and keep vendors and law enforcement accountable to the public. We take a step in this direction by making two contributions. First, our work advances gender classification benchmarking by introducing a new face dataset composed of 1270 unique individu- als that is more phenotypically balanced on")
30. Détail source à réviser : 2017) claim to determine the sexuality of Caucasian males whose profile pictures are on Facebook or dating sites. And others such as (Wu and Zhang, 2016) and Israeli based company Faception (Faception) have developed sof (Source: "2017) claim to determine the sexuality of Caucasian males whose profile pictures are on Facebook or dating sites. And others such as (Wu and Zhang, 2016) and Israeli based company Faception (Faception) have developed software that purports to deter- mine an individual’s characteristics (e.g. propen- sity towards crime, IQ, terrorism) solely from 2 Gender")
31. Détail source à réviser : detect faces (Huang et al., 2007; Kemelmacher-Shlizerman et al., 2016). Megaface, which to date is the largest publicly available set of facial images, was com- posed utilizing Head Hunter (Mathias et al., 2014) to selec (Source: "detect faces (Huang et al., 2007; Kemelmacher-Shlizerman et al., 2016). Megaface, which to date is the largest publicly available set of facial images, was com- posed utilizing Head Hunter (Mathias et al., 2014) to select one million images from the Yahoo Flicker 100M image dataset (Thomee et al., 2015; Kemelmacher-Shlizerman et al., 2016). Any sys-")
32. Détail source à réviser : lighter females, and lighter males. Due to phe- notypic imbalances in existing benchmarks, we 3 Gender Shades Figure 1: Example images and average faces from the new Pilot Parliaments Benchmark (PPB). As the examples sho (Source: "lighter females, and lighter males. Due to phe- notypic imbalances in existing benchmarks, we 3 Gender Shades Figure 1: Example images and average faces from the new Pilot Parliaments Benchmark (PPB). As the examples show, the images are constrained with relatively little variation in pose. The subjects are composed of male and female parliamentarians")
33. Détail source à réviser : 0o 60o 120o 180o 60o 0o 30o 60o 30o Figure 2: The global distribution of skin color. Most Africans have darker skin while those from Nordic countries are lighter-skinned. Image from (Encyclopedia Britannica) c©Copyright (Source: "0o 60o 120o 180o 60o 0o 30o 60o 30o Figure 2: The global distribution of skin color. Most Africans have darker skin while those from Nordic countries are lighter-skinned. Image from (Encyclopedia Britannica) c©Copyright 2012 Encyclopedia Britannica. nology (NIST) in 2015. We chose to evaluate this dataset given the government’s involvement and the")
34. Détail source à réviser : and European coun- tries. Fig. 2 shows an approximated distribu- tion of average skin types around the world. As seen in the map, African countries typically have darker-skinned individuals whereas Nordic coun- tries ten (Source: "and European coun- tries. Fig. 2 shows an approximated distribu- tion of average skin types around the world. As seen in the map, African countries typically have darker-skinned individuals whereas Nordic coun- tries tend to have lighter-skinned citizens. Col- onization and migration patterns nonetheless in- fluence the phenotypic distribution of skin")
35. Détail source à réviser : darker-skinned and female in com- parison to 21.3% in PPB. the world, the categorizations are fairly coarse. Nonetheless, the scale provides a scientifically based starting point for auditing algorithms and datasets by s (Source: "darker-skinned and female in com- parison to 21.3% in PPB. the world, the categorizations are fairly coarse. Nonetheless, the scale provides a scientifically based starting point for auditing algorithms and datasets by skin type. Gender Labels. All evaluated companies provided a “gender classification” feature that uses the binary sex labels of female")
36. Détail source à réviser : in our dataset located in each of these continents. Dataset Lighter (I,II,III) Darker (IV, V, VI) Total PPB 53.6% 681 46.4% 589 1270 IJB-A 79.6% 398 20.4% 102 500 Adience 86.2% 1892 13.8% 302 2194 Table 3: The distributi (Source: "in our dataset located in each of these continents. Dataset Lighter (I,II,III) Darker (IV, V, VI) Total PPB 53.6% 681 46.4% 589 1270 IJB-A 79.6% 398 20.4% 102 500 Adience 86.2% 1892 13.8% 302 2194 Table 3: The distributions of lighter and darker-skinned subjects (according to the Fitzpatrick clas- sification system) in PPB, IJB-A, and Adience datasets.")
37. Détail source à réviser : and darker subjects as compared to the IJB-A and Adience datasets. 4. Commercial Gender Classification Audit We evaluated 3 commercial gender classifiers. Overall, male subjects were more accurately clas- sified than fem (Source: "and darker subjects as compared to the IJB-A and Adience datasets. 4. Commercial Gender Classification Audit We evaluated 3 commercial gender classifiers. Overall, male subjects were more accurately clas- sified than female subjects replicating previous findings (Ngan et al., 2015), and lighter subjects were more accurately classified than darker in-")
38. Détail source à réviser : lacked detail and there was no mention of what training data was used. At the time of evaluation, Microsoft’s Face Detect service was described as using advanced statistical algorithms that “may not always be 100% precis (Source: "lacked detail and there was no mention of what training data was used. At the time of evaluation, Microsoft’s Face Detect service was described as using advanced statistical algorithms that “may not always be 100% precise” (Microsoft API Ref- erence). IBM Watson Visual Recognition and Face++ services were said to use deep learning- based algorithms (IBM")
39. Détail source à réviser : performance as measured by the positive predictive value (PPV), error rate (1-TPR), true positive rate (TPR), and false positive rate (FPR) of the 3 evaluated commercial classifiers on the South African subset of the PPB (Source: "performance as measured by the positive predictive value (PPV), error rate (1-TPR), true positive rate (TPR), and false positive rate (FPR) of the 3 evaluated commercial classifiers on the South African subset of the PPB dataset. Results for South Africa follow the overall trend with the highest error rates seen on darker-skinned females. the")
40. Détail source à réviser : individuals. On darker subjects, IBM achieves the worst classification accuracy with an error rate of 22.4%. This rate is nearly 7 times higher than the IBM error rate on lighter faces. Intersectional Error Rates To cond (Source: "individuals. On darker subjects, IBM achieves the worst classification accuracy with an error rate of 22.4%. This rate is nearly 7 times higher than the IBM error rate on lighter faces. Intersectional Error Rates To conduct an intersectional demographic and phenotypic analysis, the error rates for four inter- sectional groups (darker females, darker males,")
41. Détail source à réviser : composed and produced. However, darker skin alone may not be fully responsible for misclassi- fication. Instead, darker skin may be highly cor- related with facial geometries or gender display norms that were less repres (Source: "composed and produced. However, darker skin alone may not be fully responsible for misclassi- fication. Instead, darker skin may be highly cor- related with facial geometries or gender display norms that were less represented in the training data of the evaluated classifiers. 10 Gender Shades●●●● ● ●●●● ● ●●●●● ● ●●●●●● ● ● ● ● ● ● ● ● ●●●● ●● ● ● ● ●● ●● ●")
42. Détail source à réviser : though darker females are most impacted, darker males are still more misclassi- fied than lighter males for IBM and Microsoft. The most improvement is needed on darker fe- males specifically. More broadly, the error gaps (Source: "though darker females are most impacted, darker males are still more misclassi- fied than lighter males for IBM and Microsoft. The most improvement is needed on darker fe- males specifically. More broadly, the error gaps between male and female classification along with lighter and darker classification should be closed. 4.6. Accuracy Metrics Microsoft")
43. Détail source à réviser : of particular importance when doing an evaluation based on skin type. Default camera settings are often optimized to expose lighter skin better than darker skin (Roth, 2009). Underex- posed or overexposed images that pre (Source: "of particular importance when doing an evaluation based on skin type. Default camera settings are often optimized to expose lighter skin better than darker skin (Roth, 2009). Underex- posed or overexposed images that present signif- icant information loss can make accurate classi- fication challenging. With full awareness of the challenges that arise due")
44. Détail source à réviser : architectures. Because algorithmic fairness is based on differ- ent contextual assumptions and optimizations for accuracy, this work aimed to show why we need rigorous reporting on the performance metrics on which algori (Source: "architectures. Because algorithmic fairness is based on differ- ent contextual assumptions and optimizations for accuracy, this work aimed to show why we need rigorous reporting on the performance metrics on which algorithmic fairness debates center. The work focuses on increasing phenotypic and demo- graphic representation in face datasets and algo-")
45. Détail source à réviser : https://data.worldbank.org/ indicator/SG.GEN.PARL.ZS?year high desc= true. Accessed: 2017-10-06. Syafeeza Ahmad Radzi, Khalil-Hani Mohamad, Shan Sung Liew, and Rabia Bakhteri. Con- volutional neural network for face reco (Source: "https://data.worldbank.org/ indicator/SG.GEN.PARL.ZS?year high desc= true. Accessed: 2017-10-06. Syafeeza Ahmad Radzi, Khalil-Hani Mohamad, Shan Sung Liew, and Rabia Bakhteri. Con- volutional neural network for face recognition with pose and illumination variation. Interna- tional Journal of Engineering and Technology (IJET), 6(1):44–57, 2014. Julia Angwin,")
46. Détail source à réviser : gender and eth- nicity affect each other? In Informatics, Elec- tronics & Vision (ICIEV), 2012 International Conference on, pages 383–390. IEEE, 2012. Sorelle A Friedler, Carlos Scheidegger, and Suresh Venkatasubramanian (Source: "gender and eth- nicity affect each other? In Informatics, Elec- tronics & Vision (ICIEV), 2012 International Conference on, pages 383–390. IEEE, 2012. Sorelle A Friedler, Carlos Scheidegger, and Suresh Venkatasubramanian. On the (im) possibility of fairness. arXiv preprint arXiv:1609.07236, 2016. 13 Gender Shades Suranjan Ganguly, Debotosh Bhattacharjee,")
47. Détail source à réviser : 2012. Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face de- tection and recognition: Iarpa ja (Source: "2012. Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face de- tection and recognition: Iarpa janus bench- mark a. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1931–1939, 2015. Michal")
48. Détail source à réviser : recognition. In BMVC, volume 1, page 6, 2015. P Jonathon Phillips, Fang Jiang, Abhijit Narvekar, Julianne Ayyad, and Alice J O’Toole. An other-race effect for face recogni- tion algorithms. ACM Transactions on Applied Pe (Source: "recognition. In BMVC, volume 1, page 6, 2015. P Jonathon Phillips, Fang Jiang, Abhijit Narvekar, Julianne Ayyad, and Alice J O’Toole. An other-race effect for face recogni- tion algorithms. ACM Transactions on Applied Perception (TAP), 8(2):14, 2011. Alice B Popejoy and Stephanie M Fullerton. Ge- nomics is failing on diversity. Nature, 538 (7624):161, 2016.")
49. Détail source à réviser : 2005, and in the research seminar series, Department of Geog- raphy, University of Newcastle, June 2005 (Source: "2005, and in the research seminar series, Department of Geog- raphy, University of Newcastle, June 2005")
50. Détail source à réviser : 2002, 2004; Dillon & Reid, 2001; Larner & Walters, 2004) (Source: "2002, 2004; Dillon & Reid, 2001; Larner & Walters, 2004)")
51. Détail source à réviser : 2002; Valve- rde & Mopas, 2004, p (Source: "2002; Valve- rde & Mopas, 2004, p")
52. Détail source à réviser : 2004, a coalition, including the Arab-American Anti-Discrimination Committee, National Immigration Law Cen- ter, Electronic Privacy Information Center (EPIC), and American Civil Liberties Union (ACLU), wrote to the DHS e (Source: "2004, a coalition, including the Arab-American Anti-Discrimination Committee, National Immigration Law Cen- ter, Electronic Privacy Information Center (EPIC), and American Civil Liberties Union (ACLU), wrote to the DHS expressing their concern at the ‘enormous potential for error and violation of intern")
53. Détail source à réviser : 2000; Butler, 2004; De Goede, 2005), I am going to suggest here that artistic interventions in the governing of mo- bility point to important, and often neglected, forms of dissent (Source: "2000; Butler, 2004; De Goede, 2005), I am going to suggest here that artistic interventions in the governing of mo- bility point to important, and often neglected, forms of dissent")
54. Détail source à réviser : 2005, for example, the then Secretary of the Department of Homeland Security, Tom Ridge, com- pleted a number of agreements with the Dutch government to deploy biometric systems to accelerate the movement of ‘trusted tra (Source: "2005, for example, the then Secretary of the Department of Homeland Security, Tom Ridge, com- pleted a number of agreements with the Dutch government to deploy biometric systems to accelerate the movement of ‘trusted travellers’ whilst restricting the movement of higher risk groups")
55. Détail source à réviser : 2005 US ‘REAL ID’ Act is perhaps the strongest example of the move to positioning identification and credibility determina- tion, particularly of immigrants and asylum seekers, at the heart of the war on terror (Source: "2005 US ‘REAL ID’ Act is perhaps the strongest example of the move to positioning identification and credibility determina- tion, particularly of immigrants and asylum seekers, at the heart of the war on terror")
56. Détail source à réviser : L. Amoore / Political Geography 25 (2006) 336e351 authority: conferred by a raft of anti-terror legislation stitched together with the expertise of the risk managers (Source: "L. Amoore / Political Geography 25 (2006) 336e351 authority: conferred by a raft of anti-terror legislation stitched together with the expertise of the risk managers")
57. Détail source à réviser : L. Amoore / Political Geography 25 (2006) 336e351 have discussed in greater depth the problematic of resistance within the war on terror, suggest- ing that we find ourselves in ambivalent subject positions: both frequent (Source: "L. Amoore / Political Geography 25 (2006) 336e351 have discussed in greater depth the problematic of resistance within the war on terror, suggest- ing that we find ourselves in ambivalent subject positions: both frequent flier and immigrant rights campaigner, for example; or both London city commuter and anti-war protester (Amoore, 2006)")
58. Détail source à réviser : 2005 London bombings suggests renewed political fer- vour for the biometric border in the UK, with proposed new terror laws encompassing biometric immigration cards and calls for an acceleration of the US VISIT style Eur (Source: "2005 London bombings suggests renewed political fer- vour for the biometric border in the UK, with proposed new terror laws encompassing biometric immigration cards and calls for an acceleration of the US VISIT style European e-borders programme")
59. Détail source à réviser : 2005) e presumably seen as profiles of sus- picious behaviour that may have led the officer to shoot to kill (Source: "2005) e presumably seen as profiles of sus- picious behaviour that may have led the officer to shoot to kill")
60. Détail source à réviser : Biometrics: Great hope for world security or triumph of big brother? (2004, June 18). The Guardian, 17. Billions misspent in post September 11 anti terror buying. (2005, May 8). New York Times, A4. Bleiker, R. (2000). Po (Source: "Biometrics: Great hope for world security or triumph of big brother? (2004, June 18). The Guardian, 17. Billions misspent in post September 11 anti terror buying. (2005, May 8). New York Times, A4. Bleiker, R. (2000). Popular dissent, human agency and global politics. Cambridge,")
61. Détail source à réviser : 2006) 336e351 Proceedings of Machine Learning Research 81:1–15, 2018 Conference on Fairness, Accountability, and Transparency Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification∗ Joy Bu (Source: "2006) 336e351 Proceedings of Machine Learning Research 81:1–15, 2018 Conference on Fairness, Accountability, and Transparency Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification∗ Joy Buolamwini joyab@mit")
62. Détail source à réviser : 2010), biased sam- ples in AI for health care can result in treatments that do not work well for many segments of the population (Source: "2010), biased sam- ples in AI for health care can result in treatments that do not work well for many segments of the population")
63. Détail source à réviser : 2014; Bai and Ghanem, 2017), face classifica- tion (Reid et al (Source: "2014; Bai and Ghanem, 2017), face classifica- tion (Reid et al")
64. Détail source à réviser : 2017, The National Insti- tute of Standards and Technology is starting an- other challenge to spur improvement in face gen- der classification by expanding on the 2014-15 study (Source: "2017, The National Insti- tute of Standards and Technology is starting an- other challenge to spur improvement in face gen- der classification by expanding on the 2014-15 study")
65. Détail source à réviser : 2017 2015 2014 #Subjects 1270 500 2284 Avg (Source: "2017 2015 2014 #Subjects 1270 500 2284 Avg")
66. Détail source à réviser : III. Darker subjects will refer to faces labeled with a Fitzpatrick skin type of IV,V, or VI (Source: "III. Darker subjects will refer to faces labeled with a Fitzpatrick skin type of IV,V, or VI")
67. Détail source à réviser : 2015), and lighter subjects were more accurately classified than darker in- dividuals (Source: "2015), and lighter subjects were more accurately classified than darker in- dividuals")
68. Détail source à réviser : Table 5 shows that all algorithms perform worse on female and darker subjects when com- pared to their counterpart male and lighter sub- jects (Source: "Table 5 shows that all algorithms perform worse on female and darker subjects when com- pared to their counterpart male and lighter sub- jects")
69. Détail source à réviser : Is misclassification dis- tributed evenly amongst all females? Are there other factors at play? Likewise, is the misclassi- fication of darker skin uniform across gender? The intersectional error analysis that targets ge (Source: "Is misclassification dis- tributed evenly amongst all females? Are there other factors at play? Likewise, is the misclassi- fication of darker skin uniform across gender? The intersectional error analysis that targets gender classification performance on darker fe- male, lighter")
70. Détail source à réviser : 5. Conclusion We measured the accuracy of 3 commercial gen- der classification algorithms on the new Pilot Parliaments Benchmark which is balanced by gender and skin type (Source: "5. Conclusion We measured the accuracy of 3 commercial gen- der classification algorithms on the new Pilot Parliaments Benchmark which is balanced by gender and skin type")
71. Détail source à réviser : 2017 IEEE Conference on, pages 2078–2087 (Source: "2017 IEEE Conference on, pages 2078–2087")
72. Détail source à réviser : a. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1931–1939, 2015 (Source: "a. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1931–1939, 2015")
73. Détail source à réviser : Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems 29, pages (Source: "Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems 29, pages 4349–4357. Curran Associates, Inc., 2016. URL http://papers.")
74. Détail source à réviser : 2017), 2017 12th IEEE International Conference on, pages 17–24 (Source: "2017), 2017 12th IEEE International Conference on, pages 17–24")
75. Détail source à réviser : 2016) and assessing dataset diver- sity (Han and Jain, 2014), phenotypic labels are seldom used for these purposes (Source: "2016) and assessing dataset diver- sity (Han and Jain, 2014), phenotypic labels are seldom used for these purposes")
76. Détail source à réviser : In Computer Vision and Pattern Recognition Workshops (CVPRW), 2017 IEEE Conference on, pages 2078–2087 (Source: "In Computer Vision and Pattern Recognition Workshops (CVPRW), 2017 IEEE Conference on, pages 2078–2087")
77. Détail source à réviser : 2003 Accenture organized a major trans-American series of ‘citizens workshops’ on homeland security, concluding that ‘the people are the nation’s most important and untapped resource in the homeland security enterprise’ (Source: "2003 Accenture organized a major trans-American series of ‘citizens workshops’ on homeland security, concluding that ‘the people are the nation’s most important and untapped resource in the homeland security enterprise’ (Council for Excellence, 2004, p")
78. Détail source à réviser : Image. Accessed: 2018- 12-13. Proportion of seats held by women in national parliaments. https://data.worldbank.org/ indicator/SG.GEN.PARL.ZS?year high desc= true. Accessed: 2017-10-06. Syafeeza Ahmad Radzi, Khalil-Hani (Source: "Image. Accessed: 2018- 12-13. Proportion of seats held by women in national parliaments. https://data.worldbank.org/ indicator/SG.GEN.PARL.ZS?year high desc= true. Accessed: 2017-10-06. Syafeeza Ahmad Radzi, Khalil-Hani Mohamad, Shan Su")
79. Détail source à réviser : R. Garnett, editors, Advances in Neural Information Processing Systems 29, pages 4349–4357 (Source: "R. Garnett, editors, Advances in Neural Information Processing Systems 29, pages 4349–4357")
80. Détail source à réviser : De- mographic classification: Do gender and eth- nicity affect each other? In Informatics, Elec- tronics & Vision (ICIEV), 2012 International Conference on, pages 383–390. IEEE, 2012. Sorelle A Friedler, Carlos Scheidegg (Source: "De- mographic classification: Do gender and eth- nicity affect each other? In Informatics, Elec- tronics & Vision (ICIEV), 2012 International Conference on, pages 383–390. IEEE, 2012. Sorelle A Friedler, Carlos Scheidegger, and Suresh Venkatasubramanian. On the (im) possibility o")
81. Détail source à réviser : 1969; fax: þ44 0191 334 1801 (Source: "1969; fax: þ44 0191 334 1801")
82. Détail source à réviser : 2002) Introduction: homeland security/borderland insecurity At a United States House subcommittee hearing in February 2002, a panel of commercial in- formation technology experts and management consultants were asked to (Source: "2002) Introduction: homeland security/borderland insecurity At a United States House subcommittee hearing in February 2002, a panel of commercial in- formation technology experts and management consultants were asked to give technical advice on how the war on terror might be fought using risk profiling")
83. Détail source à réviser : 2006) 336e351 have discussed in greater depth the problematic of resistance within the war on terror, suggest- ing that we find ourselves in ambivalent subject positions: both frequent flier and immigrant rights campaign (Source: "2006) 336e351 have discussed in greater depth the problematic of resistance within the war on terror, suggest- ing that we find ourselves in ambivalent subject positions: both frequent flier and immigrant rights campaigner, for example; or both London city commuter and anti-war protester (Amoore, 2006)")
84. Détail source à réviser : 2016) and Israeli based company Faception (Faception) have developed software that purports to deter- mine an individual’s characteristics (e (Source: "2016) and Israeli based company Faception (Faception) have developed software that purports to deter- mine an individual’s characteristics (e")
85. Détail source à réviser : 2012 International Conference on, pages 383–390 (Source: "2012 International Conference on, pages 383–390")
86. Détail source à réviser : In Proceedings of the IEEE confer- ence on computer vision and pattern recogni- tion, pages 1701–1708, 2014 (Source: "In Proceedings of the IEEE confer- ence on computer vision and pattern recogni- tion, pages 1701–1708, 2014")
87. Détail source à réviser : 2004; De Goede, 2004; O’Malley, 2000) (Source: "2004; De Goede, 2004; O’Malley, 2000)")
88. Détail source à réviser : 2004), biometrics are parceled up, contracted out, networked and made available to multiple agencies with an anti-terror remit (Source: "2004), biometrics are parceled up, contracted out, networked and made available to multiple agencies with an anti-terror remit")
89. Détail source à réviser : 2006) 336e351 bodily experience’ (van der Ploeg, 2003; see also Thrift, 2004) (Source: "2006) 336e351 bodily experience’ (van der Ploeg, 2003; see also Thrift, 2004)")
90. Détail source à réviser : p. 269), what is ‘politically crucial’ is the necessity of thinking beyond ‘initiatory subjects and focusing on those interstitial move- ments or processes that are produced in the articulation of difference’ (Source: "p. 269), what is ‘politically crucial’ is the necessity of thinking beyond ‘initiatory subjects and focusing on those interstitial move- ments or processes that are produced in the articulation of difference’")
91. Détail source à réviser : 2006) 336e351 authority: conferred by a raft of anti-terror legislation stitched together with the expertise of the risk managers (Source: "2006) 336e351 authority: conferred by a raft of anti-terror legislation stitched together with the expertise of the risk managers")
92. Détail source à réviser : p. 145) makes an important point, then, when he argues, partially contra-Agamben, that ‘the sovereign is not simply he (or she) who first decides that there is an exception and then decides how to resolve it’ (Source: "p. 145) makes an important point, then, when he argues, partially contra-Agamben, that ‘the sovereign is not simply he (or she) who first decides that there is an exception and then decides how to resolve it’")
93. Détail source à réviser : p. 8) ‘bare life’ in which living is reduced to calculabil- ity, to the ‘life of homo sacer (Source: "p. 8) ‘bare life’ in which living is reduced to calculabil- ity, to the ‘life of homo sacer")
94. Détail source à réviser : 1988) six-point skin type scale, allowing us to benchmark the performance of gender classifica- tion algorithms by skin type (Source: "1988) six-point skin type scale, allowing us to benchmark the performance of gender classifica- tion algorithms by skin type")
95. Détail source à réviser : 2017) details the dangers and errors propa- gated by some of these aforementioned works (Source: "2017) details the dangers and errors propa- gated by some of these aforementioned works")
96. Détail source à réviser : 2015, the Adience gender and age classi- fication benchmark was released (Levi and Has- sner, 2015b) (Source: "2015, the Adience gender and age classi- fication benchmark was released (Levi and Has- sner, 2015b)")

📅 Repères chronologiques

Date	Événement
2006	Publication analysant la complexité de la vie des espèces face aux techniques biométriques
1016	Référence historique dans le contexte biométrique
2006.02	Publication sur la gouvernance des mobilités et frontières biométriques
1983	Étude sur la performance des classificateurs de genre
2000	Référence à une étude sur la discipline au travail et la gouvernance
2004	Publication sur la sécurité intérieure et la perception citoyenne de la sécurité nationale

📊 Tableaux de Synthèse

Performance des classificateurs de genre selon la couleur de peau

Type de peau	Erreur moyenne (%)
Peau claire	0,8
Peau foncée	34,7

Principaux enjeux éthiques liés à la biométrie et classification de genre

Enjeu	Description
Discrimination raciale	Biais raciaux dans la performance des algorithmes
Violation des libertés civiles	Risques de surveillance accrue et fausses identifications
Qualité des données	Impact de la diversité et de la qualité des données sur la fiabilité
Régulation	Nécessité d'une régulation pour une utilisation éthique

⚠️ Pièges & Confusions Fréquentes

1. Confusion entre identité réelle et représentation biométrique
2. Sous-estimation de la complexité de la vie des espèces face aux techniques biométriques
3. Généralisation abusive des performances des classificateurs sans considérer la démographie
4. Ignorer l'impact des biais raciaux dans la performance des algorithmes
5. Confusion entre frontières biométriques et frontières sociales
6. Négligence de la dimension éthique dans l'utilisation des technologies biométriques
7. Omettre la nécessité d'une régulation pour garantir une utilisation éthique

✅ Checklist Examen

1. Comprendre la notion de frontière biométrique et ses implications sociales
2. Analyser la performance des classificateurs selon la démographie et la couleur de peau
3. Identifier les biais raciaux dans les algorithmes de classification de genre
4. Évaluer l'impact éthique de l'utilisation des technologies biométriques
5. Reconnaître la complexité de l'identité et de la vie des espèces face aux techniques biométriques
6. Étudier la régulation nécessaire pour une utilisation responsable des technologies biométriques
7. Différencier entre vie indécidable et vie réduite à sa calculabilité
8. Analyser la relation entre biométrie, identité, et pouvoir