Bias Exploration and Mitigation in Face Presentation Attack Detection Systems

With the widespread application of face recognition systems, face presentation attack detection (PAD) plays a critical role to protect the security and credibility of the system. A growing number of researchers have investigated the biases of face recognition systems, and their results demonstrated the existence of demographic and attribute biases in recognition systems. However, the fairness of face PAD system has not attracted much attention. A key problem is the insufficient demographic and attribute annotations of face PAD data. Hence, this thesis first combines six face PAD databases: CASIA-FASD [45], REPLAYATTACK [12], MSU-MFSD [39], HKBU-MARs V1+ [30], OULU-NPU [8], WFFD [27], which consisting of print, replay, 3D mask, wax face attacks. Furthermore, identities from this combined database are manually annotated with one demographic label and six facial attribute labels. Second, this thesis explores and analyzes demographic bias and additionally facial attribute bias in face PAD methods by using this combined database. To enable the bias study, one hand-crafted feature based model LBP-MLP, and three deep learning based models: ResNet50 [23], DeepPixBis [20], and LMFD-PAD [16], are adopted. In addition to report PAD performance, a modified fairness discrepancy rate (FDR) is introduced to further determine the system fairness. The experimental results point out that deep learning based PAD models trained only on female or male group are unfairer than models trained on the fused data (including female and male). In addition, models trained on fused data and only on occlusion group indicate higher fairness than models training only on non-occlusion data. To further mitigate system bias, a modified version of PatchSwap [3], named cross-identity PatchSwap in this thesis, is introduced to enable patch substitution between identities with different gender and facial attributes. Despite the significantly improved PAD performance achieved by cross-identity PatchSwap, the FDR results also suggest that this approach is able to improve the system fairness for different gender groups when models are trained on fused data and on male data.