Mitigate Bias in Face Recognition using Skewness-Aware Reinforcement Learning (1911.10692v1)

Abstract: Racial equality is an important theme of international human rights law, but it has been largely obscured when the overall face recognition accuracy is pursued blindly. More facts indicate racial bias indeed degrades the fairness of recognition system and the error rates on non-Caucasians are usually much higher than Caucasians. To encourage fairness, we introduce the idea of adaptive margin to learn balanced performance for different races based on large margin losses. A reinforcement learning based race balance network (RL-RBN) is proposed. We formulate the process of finding the optimal margins for non-Caucasians as a Markov decision process and employ deep Q-learning to learn policies for an agent to select appropriate margin by approximating the Q-value function. Guided by the agent, the skewness of feature scatter between races can be reduced. Besides, we provide two ethnicity aware training datasets, called BUPT-Globalface and BUPT-Balancedface dataset, which can be utilized to study racial bias from both data and algorithm aspects. Extensive experiments on RFW database show that RL-RBN successfully mitigates racial bias and learns more balanced performance for different races.

Mitigating Bias in Face Recognition Through Reinforcement Learning

The paper "Mitigate Bias in Face Recognition using Skewness-Aware Reinforcement Learning" addresses a critical challenge in the deployment of face recognition (FR) systems: racial bias. This research seeks to enhance fairness in face recognition technologies, which have historically shown higher error rates for non-Caucasian individuals, thereby compromising racial equality as underscored by international human rights laws.

Racial Bias in Face Recognition Systems

The paper contextualizes its work against a backdrop of evidence indicating that FR systems tend to underperform on non-Caucasian groups. This inadequacy is attributed to both data and algorithmic biases. Commercial FR systems and state-of-the-art (SOTA) techniques often incorporate margins between class predictions to enhance accuracy, but this approach does not equitably address all racial groups due to disparate data representations and inherent recognition difficulties faced by individuals of colored skin.

Proposed Solution: Reinforcement Learning-Based Race Balance Network (RL-RBN)

The core of this research is the introduction of a sophisticated reinforcement learning (RL) framework termed Race Balance Network (RL-RBN), which aims to dynamically calibrate these margins to minimize racial bias. This paper defines the margin adjustment process as a Markov decision process and leverages deep Q-learning to construct a policy that determines race-specific margins adaptively.

Methodology

The researchers propose a multi-faceted approach:

• Ethnicity Aware Datasets: They construct two training datasets—BUPT-Globalface, which mirrors the global ethnic distribution, and BUPT-Balancedface, which ensures equal representation. These datasets are used to counterbalance inherent data biases in common FR datasets that over-represent Caucasian benchmarks.
• Deep Reinforcement Learning Application: The RL-RBN employs a deep Q-learning network to adjust model margins through skewness-aware policies, designed to minimize intra-class and inter-class distance biases across different racial groups.

Results and Implications

Extensive experiments on the Racial Faces in-the-Wild (RFW) database reveal that RL-RBN substantially mitigates racial bias within FR systems. The adaptive-margin strategy significantly reduces skewness in face recognition results, achieving more equitable performance among racial groups compared to fixed-margin methods like Cosface and Arcface.

Discussion and Future Directions

This paper contributes meaningfully to the discourse on algorithmic fairness by providing tangible methodologies for bias reduction in AI. However, it is crucial to emphasize ongoing adaptive measures alongside fixed race-balancing techniques to navigate the nuanced expectations of varied racial data dynamics continuously.

Future research may explore extending this framework beyond racial biases to encompass other demographic disparities, such as gender and age, thereby broadening the applicability of adaptive RL strategies in creating truly equitable AI systems. Additionally, the methodology's transferability to real-world applications and its long-term impact on societal trust in FR systems warrant further exploration.

In sum, the paper effectively leverages advanced machine learning techniques to address a profound ethical issue, contributing both innovative methodologies and practical solutions toward equitable technological progress.