Robust AI-Generated Video Detection via Color-Agnostic Contrastive Curriculum Learning With Color-Aware Feature Modulation

Abstract

The rapid development of video generation technologies has benefited multiple fields, but has also introduced substantial security risks and raised public concerns. A series of AI-generated video detection methods have been proposed, but they exhibit vulnerability against lossy operations, especially for color-related editing, which is widely used to improve visual appeal by common users, and even abused by attackers for anti-forensics purposes. To address this issue, we propose a robust AI-generated video detection framework via gradient-guided color-agnostic contrastive curriculum learning. Specifically, the key idea of our color-agnostic contrastive learning is to construct sample pairs adaptively and progressively, where curriculum-style iterative augmentation guided by gradient attention is developed to enrich the diversity of augmented samples and stabilize the training process. Besides, inspired by color constancy, a multi-scale color histogram pyramid is extracted with soft voting as robust auxiliary knowledge, which is injected into the visual encoder by token-wise alignment to perform feature modulation, alleviating the impact of color-related disturbance. Finally, frame-wise features are aggregated using temporal Transformer layers to identify generated videos. Extensive experiments are conducted on various generative models under common and real-world color-editing operations, where results demonstrate the superiority and flexibility of our framework compared with existing methods.