Security Analysis on Face Biometric Authentication Under Different AI-Manipulated Face

Abstract

The rapid evolution of Generative Artificial Intelligence (Gen AI) has enabled realistic facial manipulation using AI-generated images, creating serious security threats to face biometric authentication systems. Previous studies have mainly focused on deepfake detection and face recognition improvement, while limited attention has been given to the effects of AI-based facial manipulation on biometric authentication security. This study examines the vulnerability of face biometric authentication to various Gen AI-powered facial manipulation attacks. FaceForensics++ was used to create verification pairs consisting of genuine, impostor, and manipulated faces. ArcFace generated face embeddings, while cosine similarity measured identity similarity between faces. Equal Error Rate (EER) was used to determine the authentication threshold. The threshold was then applied to evaluate DeepFakes, FaceSwap, Face2Face, NeuralTextures, and FaceShifter using False Acceptance Rate (FAR), False Rejection Rate (FRR), EER, and cosine similarity scores. The results indicate different authentication behaviors among manipulation algorithms. Face2Face and NeuralTextures produced higher FAR and similarity scores, indicating greater ability to retain identity information.