A Multi-stage Bias Reduction Framework for Eye Gaze Detection

Eye gaze detection is an important component of many multimedia processing algorithms, including user authentication, online education, and medical diagnostics. Hence, it is important to ensure that gaze detection works equally well for different sections of society. For instance, if such algorithms work well for men and not women, this would amplify existing societal biases to provide more security, education, and medical functionalities to men than women. Here, we audit one of the state-of-the-art gaze detection algorithms for gender bias. Audit results suggest that the algorithm performs better for the male group than the female group, indicating a gender bias. To tackle this challenge, we propose a multi-stage bias reduction framework that considers multiple subtasks performed at different stages during the course of the gaze detection algorithm. Like many multimedia algorithms, the decisions made at each stage can impact the performance of the next algorithm stage. Hence, we have designed a framework that finds optimal algorithmic parameters to support high fairness and accuracy by holistically considering multiple stages. The results suggest that the proposed approach yields promising results in terms of fairness and accuracy, thus yielding a path toward accurate and fair eye gaze detection.