Publication

Abstract : The emergence of advanced generative models, including GANs and diffusion-based techniques, has significantly increased the spread of manipulated images, presenting a serious concern for digital forensics and content authenticity. Traditional detection approaches often fall short when it comes to handling diverse forgery types and lack the ability to provide meaningful interpretation of results. In this work, we introduce a comprehensive detection framework that combines Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) with Multi-Frequency Feature Analysis and a Self-Supervised Anomaly Detection mechanism. The CNN component is responsible for identifying fine-grained texture anomalies, while the transformer architecture captures broader contextual inconsistencies across the image. A novel Multi-Frequency Feature Module (MFFM) is designed to harness both spatial and frequency domain information, improving sensitivity to subtle alterations. To enhance the system’s adaptability, especially to unseen forgeries, a Self-Supervised Anomaly Detection (SSAD) module is integrated to model natural image distributions without the need for extensive labeled data. Experimental evaluation on multiple publicly available datasets—such as DeepFake, FaceForensics++, and GAN-generated samples—shows that the proposed method outperforms existing solutions in terms of accuracy, robustness, and generalization. Additionally, the framework supports interpretability, making it highly suitable for practical deployment in detecting synthetic media content.