GTCP: A General Traces Compensation Purifier for Enhancing the Adversarial Robustness of Deepfake Detection

Deepfake detectors achieve high accuracy on clean benchmarks but are highly vulnerable to adversarial perturbations, often dropping below 20% area under the curve (AUC) under unseen attacks. To bridge this gap without retraining existing models, we propose GTCP, a plug-and-play general trace compensation purifier that suppresses adversarial noise while compensating for essential forensic traces. GTCP operates in two stages. First, a cross-domain consistency learner (CDCL) employs supervised contrastive learning and an invariant risk minimization (IRM) inspired alignment to remove representation bias across real and forged domains, yielding a unified embedding where samples form compact, well-separated clusters under a single global decision boundary. A multi-domain forensic traces dictionary (MFTD) is then constructed over this embedding, comprising orthogonal orientation-selective “trace atoms” that precisely encode high-frequency forensic textures specific to each forgery type. Second, a collaborative reconstruction mechanism integrates the representation navigator (RN) with a novel calibration loss to dynamically select and combine the most semantically relevant trace atoms from MFTD, ensuring accurate high-frequency traces compensation. Fusing these restored details with denoised low-frequency content yields purified images that retain clean-input accuracy and substantially boost robustness against both white-box and black-box attacks. Extensive evaluations on FaceForensics++, Celeb-DF, deepfake detection challenge dataset (DFDC), and multiple detection backbones show that GTCP improves adversarial AUC by more than 60%, achieving a new state-of-the-art in robust deepfake detection.