CRDA: Adaptive Data Augmentation Based on Reinforcement Learning Improves Deepfake Detection

Existing SOTA enhancement methods for deepfake detection rely on fixed strategies and struggle to simulate the continuously evolving complexity of real-world forgeries. The CRDA framework uses a reinforcement learning agent to dynamically select forgery operations and generate adversarial samples based on the detector's current learning state, and uses causal inference to suppress spurious correlations, guiding the detector to progressively master multi-domain forgery features from simple to complex. Experiments show that the method surpasses SOTA methods on multiple cross-domain datasets and significantly improves the detector's generalization ability.