SRG-OSyNet: A Neuro-Oscillation Inspired Audio-Visual Network for Deepfake Detection

Abstract

Multimodal deepfakes pose an escalating challenge to reliable video authenticity analysis by jointly manipulating visual and auditory signals to generate highly realistic yet deceptive content. Although recent audio-visual deepfake detection methods leverage cross-modal correlations, most rely on shallow fusion strategies that inadequately capture long-range temporal dependencies and fine-grained spectral inconsistencies across modalities. Furthermore, the latent dynamical coupling that governs the coherent temporal evolution of audio and video streams is often overlooked. To address these limitations, this paper presents a Spectral-Resonance-Gated Oscillatory Synchrony Network (SRG-OSyNet) for multimodal deepfake detection, which explicitly models disruptions in audio-visual synchrony. The proposed framework incorporates a Feature-Driven Oscillatory Representation module to encode amplitude-frequency-phase interactions, a Dilated Attentive Temporal Convolutional Network to model long-term temporal dependencies, and a resonance-gating mechanism to strengthen cross-modal spectral alignment. In addition, contrastive multimodal alignment, spectral coherence regularization and oscillator stability constraints are introduced to further enhance the robustness and discriminative capacity of the learned representations. Experimental evaluations on the FakeAVCeleb and DF-TIMIT datasets demonstrate that the proposed method consistently outperforms state-of-the-art audio-visual deepfake detection approaches.