TFFNet: Texture-based Feature Fusion and Multi-scale Consistency for Deepfake Detection

Deepfake technology leverages deep learning and artificial intelligence to generate highly realistic fabricated images and videos. The ethical concerns associated with its applications in media have raised significant threats to societal trust and security. Effectively addressing the challenges posed by deepfakes requires robust legal and policy frameworks and the development of advanced technological countermeasures. Current detection methods often struggle to identify high-quality forgeries, primarily due to their inability to capture subtle texture features in shallow layers. To address these limitations, the present study introduces a novel deepfake detection framework, the Texture-based feature fusion network (TFFNet), designed to improve detection performance. TFFNet integrates texture feature fusion and multi-scale consistency, providing a more accurate and robust approach to detecting deepfakes. The texture feature fusion module strengthens the interaction between shallow and deep features across different network branches by utilizing a dual-branch structure and a cross-attention mechanism. This effectively captures fine-grained discrepancies in forged images that are typically overlooked by conventional methods. The multi-scale consistency module, in turn, applies a consistency loss function to ensure that both texture and semantic information remain consistent before and after image enhancement. Experimental evaluations conducted on the publicly available FaceForensics++ dataset demonstrate that TFFNet significantly outperforms existing detection methods in terms of accuracy, highlighting its superior ability to identify deepfakes. The code is available at https://github.com/777777 seven/TFFNet