On the "Illusion" of Gender Bias in Face Recognition: Explaining the Fairness Issue Through Non-demographic Attributes

Face recognition systems (FRS) exhibit significant accuracy differences based on the user's gender. Since such a gender gap reduces the trustworthiness of FRS, more recent efforts have tried to find the causes. However, these studies make use of manually selected, correlated, and small-sized sets of facial features to support their claims. In this work, we analyse gender bias in face recognition by successfully extending the search domain to decorrelated combinations of 40 non-demographic facial characteristics. First, we propose a toolchain to effectively decorrelate and aggregate facial attributes to enable a less-biased gender analysis on large-scale data. Second, we introduce two new fairness metrics to measure fairness with and without context. Based on these grounds, we thirdly present a novel unsupervised algorithm able to reliably identify attribute combinations that lead to vanishing bias when used as filter predicates for balanced testing datasets. The experiments show that the gender gap vanishes when images of male and female subjects share specific attributes, clearly indicating that the issue is not a question of biology but of the social definition of appearance. These findings could reshape our understanding of fairness in face biometrics and provide insights into FRS, helping to address gender bias issues.

Lights, Camera, Matching: The Role of Image Illumination in Fair Face Recognition

Facial brightness is a key image quality factor impacting face recognition accuracy differentials across demographic groups. In this work, we aim to decrease the accuracy gap between the similarity score distributions for Caucasian and African American female mated image pairs, as measured by d' between distributions. To balance brightness across demographic groups, we conduct three experiments, interpreting brightness in the face skin region either as median pixel value or as the distribution of pixel values. Balancing based on median brightness alone yields up to a 46.8% decrease in d', while balancing based on brightness distribution yields up to a 57.6% decrease. In all three cases, the similarity scores of the individual distributions improve, with mean scores maximally improving 5.9% for Caucasian females and 3.7% for African American females.

Impact of Sunglasses on One-to-Many Facial Identification Accuracy

One-to-many facial identification is documented to achieve high accuracy in the case where both the probe and the gallery are `mugshot quality' images. However, an increasing number of documented instances of wrongful arrest following one-to-many facial identification have raised questions about its accuracy. Probe images used in one-to-many facial identification are often cropped from frames of surveillance video and deviate from `mugshot quality' in various ways. This paper systematically explores how the accuracy of one-to-many facial identification is degraded by the person in the probe image choosing to wear dark sunglasses. We show that sunglasses degrade accuracy for mugshot-quality images by an amount similar to strong blur or noticeably lower resolution. Further, we demonstrate that the combination of sunglasses with blur or lower resolution results in even more pronounced loss in accuracy. These results have important implications for developing objective criteria to qualify a probe image for the level of accuracy to be expected if it used for one-to-many identification. To ameliorate the accuracy degradation caused by dark sunglasses, we show that it is possible to recover about 38% of the lost accuracy by synthetically adding sunglasses to all the gallery images, without model re-training. We also show that increasing the representation of wearing-sunglasses images in the training set can largely reduce the error rate. The image set assembled for this research will be made available to support replication and further research into this problem.

Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors

This paper studies how to synthesize face images of non-existent persons, to create a dataset that allows effective training of face recognition (FR) models. Two important goals are (1) the ability to generate a large number of distinct identities (inter-class separation) with (2) a wide variation in appearance of each identity (intra-class variation). However, existing works 1) are typically limited in how many well-separated identities can be generated and 2) either neglect or use a separate editing model for attribute augmentation. We propose Vec2Face, a holistic model that uses only a sampled vector as input and can flexibly generate and control face images and their attributes. Composed of a feature masked autoencoder and a decoder, Vec2Face is supervised by face image reconstruction and can be conveniently used in inference. Using vectors with low similarity among themselves as inputs, Vec2Face generates well-separated identities. Randomly perturbing an input identity vector within a small range allows Vec2Face to generate faces of the same identity with robust variation in face attributes. It is also possible to generate images with designated attributes by adjusting vector values with a gradient descent method. Vec2Face has efficiently synthesized as many as 300K identities with 15 million total images, whereas 60K is the largest number of identities created in the previous works. FR models trained with the generated HSFace datasets, from 10k to 300k identities, achieve state-of-the-art accuracy, from 92% to 93.52%, on five real-world test sets. For the first time, our model created using a synthetic training set achieves higher accuracy than the model created using a same-scale training set of real face images (on the CALFW test set).

Can the accuracy bias by facial hairstyle be reduced through balancing the training data?

Appearance of a face can be greatly altered by growing a beard and mustache. The facial hairstyles in a pair of images can cause marked changes to the impostor distribution and the genuine distribution. Also, different distributions of facial hairstyle across demographics could cause a false impression of relative accuracy across demographics. We first show that, even though larger training sets boost the recognition accuracy on all facial hairstyles, accuracy variations caused by facial hairstyles persist regardless of the size of the training set. Then, we analyze the impact of having different fractions of the training data represent facial hairstyles. We created balanced training sets using a set of identities available in Webface42M that both have clean-shaven and facial hair images. We find that, even when a face recognition model is trained with a balanced clean-shaven / facial hair training set, accuracy variation on the test data does not diminish. Next, data augmentation is employed to further investigate the effect of facial hair distribution in training data by manipulating facial hair pixels with the help of facial landmark points and a facial hair segmentation model. Our results show facial hair causes an accuracy gap between clean-shaven and facial hair images, and this impact can be significantly different between African-Americans and Caucasians.

What is a Goldilocks Face Verification Test Set?

Face Recognition models are commonly trained with web-scraped datasets containing millions of images and evaluated on test sets emphasizing pose, age and mixed attributes. With train and test sets both assembled from web-scraped images, it is critical to ensure disjoint sets of identities between train and test sets. However, existing train and test sets have not considered this. Moreover, as accuracy levels become saturated, such as LFW $>99.8\%$, more challenging test sets are needed. We show that current train and test sets are generally not identity- or even image-disjoint, and that this results in an optimistic bias in the estimated accuracy. In addition, we show that identity-disjoint folds are important in the 10-fold cross-validation estimate of test accuracy. To better support continued advances in face recognition, we introduce two "Goldilocks" test sets, Hadrian and Eclipse. The former emphasizes challenging facial hairstyles and latter emphasizes challenging over- and under-exposure conditions. Images in both datasets are from a large, controlled-acquisition (not web-scraped) dataset, so they are identity- and image-disjoint with all popular training sets. Accuracy for these new test sets generally falls below that observed on LFW, CPLFW, CALFW, CFP-FP and AgeDB-30, showing that these datasets represent important dimensions for improvement of face recognition. The datasets are available at: \url{https://github.com/HaiyuWu/SOTA-Face-Recognition-Train-and-Test}

Identity Overlap Between Face Recognition Train/Test Data: Causing Optimistic Bias in Accuracy Measurement

A fundamental tenet of pattern recognition is that overlap between training and testing sets causes an optimistic accuracy estimate. Deep CNNs for face recognition are trained for N-way classification of the identities in the training set. Accuracy is commonly estimated as average 10-fold classification accuracy on image pairs from test sets such as LFW, CALFW, CPLFW, CFP-FP and AgeDB-30. Because train and test sets have been independently assembled, images and identities in any given test set may also be present in any given training set. In particular, our experiments reveal a surprising degree of identity and image overlap between the LFW family of test sets and the MS1MV2 training set. Our experiments also reveal identity label noise in MS1MV2. We compare accuracy achieved with same-size MS1MV2 subsets that are identity-disjoint and not identity-disjoint with LFW, to reveal the size of the optimistic bias. Using more challenging test sets from the LFW family, we find that the size of the optimistic bias is larger for more challenging test sets. Our results highlight the lack of and the need for identity-disjoint train and test methodology in face recognition research.

CRAFT: Contextual Re-Activation of Filters for face recognition Training

The first layer of a deep CNN backbone applies filters to an image to extract the basic features available to later layers. During training, some filters may go inactive, mean ing all weights in the filter approach zero. An inactive fil ter in the final model represents a missed opportunity to extract a useful feature. This phenomenon is especially prevalent in specialized CNNs such as for face recogni tion (as opposed to, e.g., ImageNet). For example, in one the most widely face recognition model (ArcFace), about half of the convolution filters in the first layer are inactive. We propose a novel approach designed and tested specif ically for face recognition networks, known as "CRAFT: Contextual Re-Activation of Filters for Face Recognition Training". CRAFT identifies inactive filters during training and reinitializes them based on the context of strong filters at that stage in training. We show that CRAFT reduces fraction of inactive filters from 44% to 32% on average and discovers filter patterns not found by standard training. Compared to standard training without reactivation, CRAFT demonstrates enhanced model accuracy on standard face-recognition benchmark datasets including AgeDB-30, CPLFW, LFW, CALFW, and CFP-FP, as well as on more challenging datasets like IJBB and IJBC.

LogicNet: A Logical Consistency Embedded Face Attribute Learning Network

Ensuring logical consistency in predictions is a crucial yet overlooked aspect in multi-attribute classification. We explore the potential reasons for this oversight and introduce two pressing challenges to the field: 1) How can we ensure that a model, when trained with data checked for logical consistency, yields predictions that are logically consistent? 2) How can we achieve the same with data that hasn't undergone logical consistency checks? Minimizing manual effort is also essential for enhancing automation. To address these challenges, we introduce two datasets, FH41K and CelebA-logic, and propose LogicNet, an adversarial training framework that learns the logical relationships between attributes. Accuracy of LogicNet surpasses that of the next-best approach by 23.05%, 9.96%, and 1.71% on FH37K, FH41K, and CelebA-logic, respectively. In real-world case analysis, our approach can achieve a reduction of more than 50% in the average number of failed cases compared to other methods.

Our Deep CNN Face Matchers Have Developed Achromatopsia

Modern deep CNN face matchers are trained on datasets containing color images. We show that such matchers achieve essentially the same accuracy on the grayscale or the color version of a set of test images. We then consider possible causes for deep CNN face matchers ``not seeing color''. Popular web-scraped face datasets actually have 30 to 60\% of their identities with one or more grayscale images. We analyze whether this grayscale element in the training set impacts the accuracy achieved, and conclude that it does not. Further, we show that even with a 100\% grayscale training set, comparable accuracy is achieved on color or grayscale test images. Then we show that the skin region of an individual's images in a web-scraped training set exhibit significant variation in their mapping to color space. This suggests that color, at least for web-scraped, in-the-wild face datasets, carries limited identity-related information for training state-of-the-art matchers. Finally, we verify that comparable accuracy is achieved from training using single-channel grayscale images, implying that a larger dataset can be used within the same memory limit, with a less computationally intensive early layer.