Fairer Face Recognition Through Individual Score Normalisation

With their widespread use in various fields, including security, finance, and transportation, face recognition systems have become an essential tool for verifying individuals’ identities. These systems offer ease of use and accuracies that surpass those of humans. Nevertheless, in recent work, face recognition systems have been shown to be biased toward different demographic attributes such as age, ethnicity, and gender. That is, for specific subgroups of these attributes, the proportion of errors produced by the systems is significantly higher than for other subgroups. This is especially concerning as these systems are widely deployed in critical areas such as forensics in court, where their decisions can have a significant impact on the people. Thus, a solution is needed that mitigates this bias while still providing good system accuracy. In previous work, methods mainly focus on learning less biased representations of faces to achieve fairer systems. However, these methods are often hardly integrable into existing face recognition systems and reduce the overall recognition performance while mitigating the bias. This thesis proposes a method for Fair Adaptation Through Local Optimal (Threshold) Normalisation called FALCON, which can be seamlessly integrated as a post-processing method, works in an unsupervised manner (i.e. no demographic labels are required for the training data), and significantly increases both fairness and accuracy of face recognition systems. FALCON processes each face image individually and reduces bias by treating individuals with similar characteristics similarly. The proposed method is tested, evaluated, and compared with four other advanced post-processing methods across four distinct face recognition models and three publicly available datasets. The results demonstrate that FALCON significantly enhances the fairness and accuracy of face recognition systems, providing a lightweight solution that can be easily integrated into existing systems. Unlike other post-process methods, FALCON provides a parameter to adjust the trade-off between fairness and accuracy. The proposed method offers a well-working solution to address the bias in face recognition systems, making these systems fairer for all users, regardless of their demographics.