Abstract
Deepfake detection systems have achieved impressive accuracy on conventional forged images; however, they remain vulnerable to anti-forensic or adversarial samples deliberately crafted to evade detection. Such samples introduce imperceptible perturbations that conceal forgery artifacts, causing traditional binary classifiers-trained solely on real and forged data-to misclassify them as authentic. In this paper, we address this challenge by proposing a multi-channel feature extraction framework combined with a three-class classification strategy. Specifically, one channel focuses on extracting identity-preserving facial representations to capture inconsistencies in personal identity traits, while additional channels extract complementary spatial and frequency domain features to detect subtle forgery traces. These multi-channel features are fused and fed into a three-class detector capable of distinguishing real, forged, and anti-forensic samples. Experimental results on datasets incorporating adversarial deepfakes demonstrate that our method substantially improves robustness against anti-forensic attacks while maintaining high accuracy on conventional deepfake detection tasks.