A Comprehensive Study on Face Recognition Biases Beyond Demographics

Terhörst, Philipp; Kolf, Jan Niklas; Huber, Marco F.; Kirchbuchner, Florian; Damer, Naser; Morales, Aythami; Fiérrez, Julián; Kuijper, Arjan

doi:10.1109/tts.2021.3111823

Cited by 76 publications

(26 citation statements)

References 77 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For APS, we varied the adversarial network model size from one up to three layers, and layer size (including the output dimension H) from 32 to 128 hidden units. Overall EER ranged from 1.09% to 1.15% for APS (10) and from 1.12% to 1.16% for APS (11); the best result was obtained with one hidden layer of 64 units and 128 units, respectively. For PL and PW, K-means clustering is performed at the beginning of each epoch with K = 8, 32, and 128.…”

Section: Resultsmentioning

confidence: 99%

“…Here f φ (x a j ) ∈ R H is the adversarial network, and H is its output embedding dimension, a hyperparameter. We use the exponential of cosine similarity in (11) with an assumption that it is normally distributed.…”

Section: Accumulated Pairwise Similarity (Aps)mentioning

confidence: 99%

“…To address the fairness problem, many researches have been focusing on identifying the biases in the training data [5,6]. These include applications to face recognition [10,11], image classification [12], speech recognition [13], natural language processing [14,15], and classification [16,17]. For speaker verifica-tion, Fenu et al [18] showed how data balancing in training can improve fairness.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Adversarial Reweighting for Speaker Verification Fairness

Jin¹,

Ju²,

Chen³

et al. 2022

Preprint

View full text Add to dashboard Cite

We address performance fairness for speaker verification using the adversarial reweighting (ARW) method. ARW is reformulated for speaker verification with metric learning, and shown to improve results across different subgroups of gender and nationality, without requiring annotation of subgroups in the training data. An adversarial network learns a weight for each training sample in the batch so that the main learner is forced to focus on poorly performing instances. Using a min-max optimization algorithm, this method improves overall speaker verification fairness. We present three different ARW formulations: accumulated pairwise similarity, pseudo-labeling, and pairwise weighting, and measure their performance in terms of equal error rate (EER) on the VoxCeleb corpus. Results show that the pairwise weighting method can achieve 1.08% overall EER, 1.25% for male and 0.67% for female speakers, with relative EER reductions of 7.7%, 10.1% and 3.0%, respectively. For nationality subgroups, the proposed algorithm showed 1.04% EER for US speakers, 0.76% for UK speakers, and 1.22% for all others. The absolute EER gap between gender groups was reduced from 0.70% to 0.58%, while the standard deviation over nationality groups decreased from 0.21 to 0.19.

show abstract

Section: Resultsmentioning

confidence: 99%

Section: Accumulated Pairwise Similarity (Aps)mentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Adversarial Reweighting for Speaker Verification Fairness

Jin¹,

Ju²,

Chen³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…Generally, biometric systems perform differently on different groups of users [25,3,8]. In [27,5], several user groups have been characterized depending on their effect on the biometric system.…”

Section: The Dodding Zoo and Masterfacesmentioning

confidence: 99%

On the (Limited) Generalization of MasterFace Attacks and Its Relation to the Capacity of Face Representations

Terhörst¹,

Bierbaum²,

Huber³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

A MasterFace is a face image that can successfully match against a large portion of the population. Since their generation does not require access to the information of the enrolled subjects, MasterFace attacks represent a potential security risk for widely-used face recognition systems. Previous works proposed methods for generating such images and demonstrated that these attacks can strongly compromise face recognition. However, previous works followed evaluation settings consisting of older recognition models, limited cross-dataset and cross-model evaluations, and the use of low-scale testing data. This makes it hard to state the generalizability of these attacks. In this work, we comprehensively analyse the generalizability of MasterFace attacks in empirical and theoretical investigations. The empirical investigations include the use of six state-of-the-art FR models, cross-dataset and cross-model evaluation protocols, and utilizing testing datasets of significantly higher size and variance. The results indicate a low generalizability when MasterFaces are training on a different face recognition model than the one used for testing. In these cases, the attack performance is similar to zero-effort imposter attacks. In the theoretical investigations, we define and estimate the face capacity and the maximum MasterFace coverage under the assumption that identities in the face space are well separated. The current trend of increasing the fairness and generalizability in face recognition indicates that the vulnerability of future systems might further decrease. Future works might analyse the utility of MasterFaces for understanding and enhancing the robustness of face recognition models.

show abstract

“…The rapid adoption and deployment of facial recognition technology in recent years has effected a startling discovery: the accuracy of facial recognition may vary -significantly -based on the color of one's skin. Both the biometrics community and the public were shocked by this revelation, which has since been widely discussed in news stories [25,30,39,43] and research papers [10,15,18,21,29,42].…”

Section: Introductionmentioning

confidence: 99%