HARD-Xception

A Hybrid Adversarially Robust Deepfake Detection Framework Using Frequency Decomposition and Feature Consistency Learning

Abstract

Deepfake detection systems achieve strong performance on clean datasets but remian highly vulnerable to adversarial perturbations and cross-dataset distribution shifts. We present HARD-Xception, a hybrid adversarially robust deepfake detection framework designed to improve robustness under these conditions. Input face images are decomposed into disjoint frequency bands using the Discrete Cosine Transform, and each band is processed by an independent Xception-based branch to learn complementary forensic cues. The resulting embeddings are fused for classification. To improve robustness, we incorporate projected gradient descent-based adversarial training and enforce feature-level consistency between clean and adversarial representations using maximum mean discrepancy and center loss regularization. Preliminary experiments on RealVsFake and FaceForensics++ demonstrate meaningful discriminative performance under clean evaluation and improved recall and AUC under adversarial and cross-dataset settings. These results highlight the importance of frequency-aware representations and feature stability for robust deepfake detection.