Intra-identity PatchSwap: On the Generalizability of Face Presentation Attack Detection
With the widespread deployment of face recognition systems, face presentation attack detection (PAD) plays an essential role in mitigating their vulnerabilities. Face PAD is employed before the identification system to detect if the presented face is an attack. However, most of the existing face PAD methods tend to overfit on the training data and fail to generalize well on unknown attacks in a real-world scenario. The main reason for such poor generalizability is that existing face PAD datasets are limited in quantity and diversity. Moreover, recent PAD works leverage pixel-wise supervision strategy and show great progress in face PAD performance. Nevertheless, obtaining accurate pixelwise labels is a challenging task. To alleviate these issues, we propose the plug-n-play PatchSwap approach in this thesis. The proposed PatchSwap method maximizes limited data utilization and generates more challenging bonafide/attack samples and partial attacks by swapping patches between training data by a well-designed strategy. Meanwhile, their pixel-wise labels are correspondingly updated. As a result, the augmented training samples contain more complex attack patterns, benefiting robust feature learning. Furthermore, we demonstrated the proposed PatchSwap method combined with three prevailing backbones: ResNet, DenseNet, and MixFaceNet. The extensive experiments were performed on four benchmark datasets under both intra-dataset and cross-dataset scenarios. We also conducted several detailed ablation studies to explore the effect of patch types, selected candidate identity, and the probabilities controlling the swapping process. The experimental results show that the proposed PatchSwap approach achieved significant performance improvement. For example, the ACER value on the most challenging Protocol-4 of Oulu-NPU  decreased from 20.51% achieved by DenseNet baseline to 3.41% by DenseNet-PatchSwap.
Darmstadt, TU, Master Thesis, 2022