Multi-Domain Deepfake Detection Framework Using Transformer-Based Cross-Attention Fusion

Author : Sanjaya Shankar Tripathy, Tarush Anand, Chetan Singh, Rounit Kumar Sinha

Abstract : Modern generative models such as style-based Generative Adversarial Networks (StyleGAN) synthesize facial images realistic enough that single-cue deepfake detectors are no longer reliable. This paper presents a quad-stream deepfake detection framework that jointly analyses a facial image from spatial, frequency, noise-residual, and statistical forensic perspectives. Stream 1 uses a ViT-Base/16 or Swin-Transformer backbone with an auxiliary mask-supervision branch for global context and forgery localisation. Stream 2 uses a modern CNN backbone (EfficientNet-V2-S / ConvNeXt) for spatial texture cues. Stream 3 forms a five-channel Learned Noise Pattern (LNP) fingerprint from deep denoising residuals fused with FFT amplitude and phase spectra. Stream 4 fits a Generalised Gaussian Distribution to the 63 AC coefficients of an 8×8 block DCT to build a 126-D statistical fingerprint. All four 256-D stream embeddings are combined through a Gated Cross-Attention Fusion module that adaptively re-weights each stream. Trained and evaluated on the 140k Real and Fake Faces dataset (70,000 real FFHQ and 70,000 StyleGAN images), the proposed system attains a test AUC-ROC of 0.9371 and an overall accuracy of 89.18% on 20,000 unseen samples, with validation AUC reaching 0.9905 after 15 epochs.

Keywords : Deepfake detection, StyleGAN, Vision Transformer, frequency-domain analysis, noise residual fingerprinting, multi-stream fusion, gated cross-attention

Conference Name : International Conference on Computer Vision and Artificial Intelligence (ICCVAI - 26)

Conference Place : Visakhapatnam, India

Conference Date : 29th Aug 2026