Hybrid Spatial-Frequency Deepfake Detection with Adaptive Fusion and Domain-Adversarial Learning

---

Résumé

The rapid proliferation of AI-generated deepfakes threatens digital security and information integrity. Existing detectors suffer from poor cross-dataset generalization, degrading AUC on unseen data. This paper proposes a twobranch hybrid architecture combining spatial and frequency analysis via two parallel EfficientNet-B0 backbones fused by an adaptive attention module. A Domain-Adversarial Neural Network (DANN) with Gradient Reversal Layer (GRL) enforces domain-invariant feature learning. Frame-level predictions are aggregated via logistic regression over 12 temporal descriptors. Under a multi-source held-out evaluation protocol on DFDC, FaceForensics++, and Celeb-DF (training on all three, testing on held-out splits), the compact 8.8M-parameter system achieves video-level AUC = 0.9928 (±0.002, 95% CI) with consistent performance across held-out test splits (generalization gap of -1.2%). A web demonstration is available at https://github.com/ Enochbi/Deepfake-detection.

Mots-clés

deepfake detection, video forgery detection, spatial-frequency fusion, adaptive attention, domain adversarial neural networks, gradient reversal layer, cross-dataset generalization, EfficientNetB0, DFDC, FaceForensics++, Celeb-DF