A Principal Component Analysis-Based Approach for Single Morphing Attack Detection

This paper proposes an explicit method for single face image morphing attack detection, using an RGB decomposition based on Principal Component Analysis from texture patterns. Handcrafted detection algorithms can be advantageous over deep learning-based methods as they constitute increased explainability, showcased in this work by visualizing relevant face areas for morphing attack detection. Such information can be relevant for deployed systems in real-world scenarios with humans in the loop. The morphing detection capability of the proposed method is evaluated extensively across three datasets and six morphing algorithms in single, cross-dataset and cross-morphed scenarios and compared to a fine-tuned MobileNetV2 architecture. The results show how single image morphing attack detection remains challenging, especially in cross-domain scenarios involving realistic diversity of morphing algorithms, including StyleGAN-based approaches. In such conditions, the proposed method can be as good or even better than the evaluated MobileNetV2 approach.