Dual-stream deepfake detector in PyTorch: ConvNeXt-Small spatial stream + ResSE-Spectral Tower (SRM/Bayar/FFT) frequency stream with SNR-adaptive gating, fused into a Bi-GRU video head. Zero-leakage identity splits, calibrated 3-zone thresholds, Grad-CAM diagnostics. Video AUC 0.872, LOTO macro-AUC 0.897, 60.9 FPS on T4.
opencv computer-vision grad-cam pytorch pytest networkx graph-partitioning torchvision model-interpretability deepfake-detection streamlit convnext huggingface-spaces dual-stream gated-fusion fft-spectrum spatial-rich-model bayar-stamm temperature-calibration loss-masking
-
Updated
Sep 18, 2026 - Python