Using DCGANs and HOG + Patch-Based CNN for Face Spoofing Mitigation

The use of biometric authentication systems is replacing traditional knowledge-based and token-based systems. With the proliferation of inexpensive cameras and scanners, along with improved recognition algorithms and protocols, the distinctiveness of biometrics ensures the efficacy of authentication systems. Although the use of biometrics is making systems more efficient, there are still security vulnerabilities that can be exploited by biometric spoofing attacks. Most recently, the use of artificial intelligence, such as deepfakes, have been the popular source of quality spoofing attacks. Deepfakes leverage the power of machine learning techniques and artificial intelligence to manipulate or generate visual content with a higher potential to deceive. In this work, we explore spoofing mitigation techniques by using the DCGAN framework to generate high quality synthetic images, using face biometric samples. We then proposed a HOG + Patch-based CNN structure for spoofing mitigation on the generated spoofing datasets. Our proposed CNN model outperforms notable VGG-16 and ResNet-50 models in classification and verification accuracies on our spoofing dataset.