Towards Unified Face Verification Against Quality Variations and Deepfake

Face verification aims to provide reliable biological verification results for applications in areas such as security biometric identification, judicial analysis, and digital identity platforms. However, low-quality and malicious deepfake samples caused by complex collection environments may lead to erroneous outcomes in face verification. This paper introduces a unified system for reliable face verification designed to counteract deepfakebased identity spoofing while ensuring input quality reliability. The system processes an input pair of face images or videos through a multi-module pipeline. Specifically, a face detection module crops and aligns firstly the facial regions of input sample pairs. Then, in order to measure and reject low-quality samples, we design a dual metric-driven quality assessment module via class-centric deviation and embedding uncertainty fusion. Sequentially, the deepfake detection module will filter out the input of illegally synthetic face-swapping samples for anti-spoofing. For genuine samples, a quality-aware verification module trained by self-distillation is proposed for the final face verification. Herein, during the training process, low-quality samples are guided by high-quality samples to converge toward class centroids by minimizing the Wasserstein distance between their feature distributions, thereby enhancing model accuracy without increasing spatial complexity. Extensive experiments verify the reliability of our system.