Fairness Under Cover: Evaluating the Impact of Occlusions on Demographic Bias in Facial Recognition (2408.10175v1)

Abstract: This study investigates the effects of occlusions on the fairness of face recognition systems, particularly focusing on demographic biases. Using the Racial Faces in the Wild (RFW) dataset and synthetically added realistic occlusions, we evaluate their effect on the performance of face recognition models trained on the BUPT-Balanced and BUPT-GlobalFace datasets. We note increases in the dispersion of FMR, FNMR, and accuracy alongside decreases in fairness according to Equilized Odds, Demographic Parity, STD of Accuracy, and Fairness Discrepancy Rate. Additionally, we utilize a pixel attribution method to understand the importance of occlusions in model predictions, proposing a new metric, Face Occlusion Impact Ratio (FOIR), that quantifies the extent to which occlusions affect model performance across different demographic groups. Our results indicate that occlusions exacerbate existing demographic biases, with models placing higher importance on occlusions in an unequal fashion, particularly affecting African individuals more severely.

• The paper evaluates facial recognition bias under occlusions using metrics like FOIR, showing decreased accuracy and increased error dispersion across demographics.
• Research highlights that facial occlusions disproportionately degrade facial recognition performance for African individuals.
• The findings emphasize the need to evaluate facial recognition fairness under real-world conditions like occlusion and suggest training models resilient to these challenges.

An Evaluation of Demographic Bias in Occluded Face Recognition Systems

The academic paper titled "Fairness Under Cover: Evaluating the Impact of Occlusions on Demographic Bias in Facial Recognition" focuses on the impact of occlusions on the fairness of face recognition systems, particularly in relation to demographic biases. This paper utilizes facial recognition models trained on the BUPT-Balanced and BUPT-GlobalFace datasets and evaluates them against the Racial Faces in the Wild (RFW) dataset with synthetically added occlusions. The objective is to ascertain the effect of occlusions on model performance and whether these exacerbate existing biases across different demographic groups.

The research methodology involves an experimental design that assesses model performance with and without occlusions. The occlusions are added following protocols that reflect realistic conditions such as masks and sunglasses. Performance is measured in terms of accuracy, False Match Rate (FMR), False Non-Match Rate (FNMR), and various fairness metrics including Equalized Odds and Demographic Parity. One of the novel contributions of this paper is the introduction of the Face Occlusion Impact Ratio (FOIR), a metric designed to quantify the impact of occlusions on model predictions across different demographics.

Key Findings

1. Performance and Fairness Metrics: The presence of occlusions leads to an overall increase in errors, with a significant drop in accuracy and biased error distribution across demographic groups. This is reflected in increased standard deviation (STD) of accuracies and worse Demographic Parity (DP) and Equalized Odds (EO). Interestingly, metrics dependent on ratios of error rates like Skewed Error Ratio (SER), Inequity Rate (IR), and GARBE yield apparently fairer outcomes due to the large absolute increase in error rates, although the dispersion of error rates across demographics also increases.
2. Impact on Different Ethnicities: The research highlights that occlusions have a disproportionate impact on African individuals, who experience higher performance degradation compared to other groups. This is demonstrated by significant differences in the FOIR metrics, especially in False Non-Match Rate (FNM) scenarios where genuine pairs are misclassified.
3. Pixel Attribution Analysis: Utilizing explainable AI tools, the paper examines the distribution of pixel importance concerning occlusions. For African individuals, occluded regions have a more pronounced influence on erroneous predictions, indicating unequal treatment by the models across ethnicities.

Implications and Future Directions

The findings of this paper underline the necessity of examining fairness in face recognition systems under conditions that reflect real-world scenarios, such as facial occlusions. The exacerbation of demographic biases due to occlusions poses significant challenges for deploying such models in sensitive applications. The paper suggests a focus on training models that are resilient to occlusions and implementing fairness evaluations that consider occluded conditions.

Future work could explore refining fairness metrics to better accommodate the nuances of high-error scenarios. Developing more robust models that account for occlusions during training may also help in mitigating bias. Additionally, examining the inner workings of models through explainable AI and refining metrics like the FOIR could offer deeper insights into addressing demographic biases in occluded environments. This effort is crucial for ensuring that face recognition technology is deployed ethically and equitably across diverse populations.