Post-Comparison Mitigation of Demographic Bias in Face Recognition Using Fair Score Normalization

Current face recognition systems achieve high progress on several benchmark\ntests. Despite this progress, recent works showed that these systems are\nstrongly biased against demographic sub-groups. Consequently, an easily\nintegrable solution is needed to reduce the discriminatory effect of these\nbiased systems. Previous work mainly focused on learning less biased face\nrepresentations, which comes at the cost of a strongly degraded overall\nrecognition performance. In this work, we propose a novel unsupervised fair\nscore normalization approach that is specifically designed to reduce the effect\nof bias in face recognition and subsequently lead to a significant overall\nperformance boost. Our hypothesis is built on the notation of individual\nfairness by designing a normalization approach that leads to treating similar\nindividuals similarly. Experiments were conducted on three publicly available\ndatasets captured under controlled and in-the-wild circumstances. Results\ndemonstrate that our solution reduces demographic biases, e.g. by up to 82.7%\nin the case when gender is considered. Moreover, it mitigates the bias more\nconsistently than existing works. In contrast to previous works, our fair\nnormalization approach enhances the overall performance by up to 53.2% at false\nmatch rate of 0.001 and up to 82.9% at a false match rate of 0.00001.\nAdditionally, it is easily integrable into existing recognition systems and not\nlimited to face biometrics.\n