Deepfake Detection via Spatial-Frequency Attention Network

Artificial Intelligence Generated Content (AIGC) makes creating realistic synthetic media a breeze. Facial forgery technologies threaten consumer electronics, such as smartphones, intelligent payment systems, and smart home Internet of Things (IoT) devices. These threats undermine identity authentication security and compromise user privacy. To address this problem, this paper introduces a wavelet-based Spatial and Frequency Domain Attention Network (SFANet) for deepfake detection. The SFANet comprises a wavelet transform-based feature decoupling module, a dual attention module, and a cross-modality fusion module. The spatial features are first decomposed into high and low-frequency components by wavelet transform. Subsequently, a dual attention mechanism is built to adjust frequency-domain weights dynamically. Finally, a cross-modality fusion module is introduced to integrate spatial and frequency-domain features for enhanced detection. Experimental results on three benchmark datasets, namely FaceForensics++, WildDeepfake, and Celeb-DF V2, prove that SFANet achieves superior detection performance compared to state-of-the-art models. This research provides a powerful tool to combat AI-driven misinformation and advances the field of synthetic content detection in consumer electronics. The implementation of SFANet will be publicly available at.