A Multi-Attribute CNN Framework for Deepfake Forensics

The challenge of deepfake detection is severely hampered by a “Generalization Crisis,” where detectors fail when encountering unseen generation techniques or data compression levels due to over-reliance on easily suppressed, domain-specific artifacts. To address this, we propose the Multi-Attribute Convolutional Neural Network (MA-CNN) framework, specifically designed for advanced deepfake forensics. The MA-CNN utilizes a highly modular, three-stream architecture that achieves simultaneous and decoupled extraction of three orthogonal forensic attributes: Frequency Residues (AF ), Low-Level Steganalysis Traces (AS ), and Physiological Inconsistencies (AP ). The core innovation is the Dynamic Fusion Module (DFM), which employs Factorized Bilinear Pooling (FBP) and channel-wise attention to capture complex, second-order interdependencies between arti- facts, maximizing the discriminative power. Rigorous evaluation under a demanding cross-dataset generalization protocol (trained on FaceForensics++ and tested on Celeb-DF v2 and DFDC) demonstrates that the MA-CNN achieves superior performance compared to state-of-the-art baselines. This study validates the hypothesis that explicit engineering of multi-attribute feature extractors yields a robust, domain-invariant detection capability, confirmed by interpretability analysis showing the model prioritizes low-level forensic evidence.