Authors - Priyanka Khalate, Satish S. Banait, Chandrakant Kokane, Dnyanada Shinde, Madhumati Pol, Pravinkumar M. Sonsare Abstract - The emerging use of digital deepfake technology is creating a myriad of obstacles in verifying the authenticity of digital media. Most of today’s detection methods yield satisfactory results when applied to clean samples of content, however, they are still susceptible to adversarial perturbations specifically created to bypass these detection methods. The current research paper introduces DC-DAFDN, a dual-stream architecture for detecting fraudulent digital content, which fuses frequency-domain analysis using the Discrete Cosine Transform (DCT) with Space-Attention Mechanisms. The current architecture uses adversarial training to develop more robust features. The proposed model uses EfficientNet-B4 as a backbone, augmented with Spatial Reduction Attention Blocks and Forged Fea tures Attention Modules to detect manipulation artifacts in the spatial domain, while the parallel DCT stream analyzes inconsistencies in the frequency-domain. Through an adversarial training procedure using Fast Gradient Sign Method (FGSM)-induced adversarial perturbations, the model learns robust feature sets that are resistant to evasion attacks. When evaluated on Face-Forensics++ dataset, DC-DAFDN significantly improves upon the original Dual Attention for Deepfake Detection Network (DAFDN) in terms of adversarial robustness. When attacked with large adversarial perturbations (e.g., FGSM with ϵ ranging from 0.1 to 0.25), the DC-DAFDN architecture maintained greater than average accu racy enhancements from +2.74% up to +3.61%, for an average accuracy increase of +3.36%, for the tested att, from all strengths. Our findings suggest that fusing frequency-domain analysis with adversarial training provides measurable improvement in the model’s robustness to adversarial attacks and simultaneously preserves the detection capabilities of the dual-attention method.
Open Zoom