ABSTRACT To address the critical security challenges in AI face‐swap detection—particularly the trade‐off between model lightweighting and robustness enhancement, coupled with insufficient generalization against adversarial attacks—this study proposes a security‐oriented multidimensional feature fusion framework for real‐time forgery identification. First, five complementary feature dimensions—mirror similarity, geometric asymmetry, frequency‐domain anomalies, image quality artifacts, and DeepFace confidence deviation—are integrated into a security‐oriented feature set. This set targets inherent vulnerabilities in face‐swap forgeries, enabling multiperspective identification of tampered faces and reducing the risk of missed detections or false positives in security‐sensitive scenarios; Second, Design of an adaptive security threshold mechanism based on dynamic confidence intervals, which continuously optimizes decision boundaries through online statistical modeling to strengthen defense against unseen attacks; Finally, Implementation of a lightweight random forest ensemble learning paradigm that delivers millisecond‐level inference while maintaining interpretability. Rigorous evaluations on benchmark datasets (FaceForensics++ and Celeb‐DF) demonstrate superior performance over conventional ResNet50 approaches, achieving 95.1% average accuracy and 93.4% F1‐score with merely 68 ms per‐frame processing latency. Notably, our framework exhibits exceptional security resilience in countering advanced deepfake assaults.
Zhang et al. (Tue,) studied this question.