Evaluating Content Consistency in Encoded Images and Videos for Forgery Detection

Detecting forged images and Deepfake videos remains a pressing challenge. Most existing detection methods rely on datasets that contain specific types of tampering. However, as forgery techniques continue to evolve, new forms of manipulation emerge and may undermine these approaches. Furthermore, media content often undergoes compression during dissemination, which can obscure manipulation traces and reduce detection accuracy. This study proposes a forgery detection scheme that leverages block-level consistency. A source-related feature extractor is employed, followed by a Siamese network to classify the compression levels. For image tampering detection, the heatmap is examined to determine manipulated regions. In Deepfake video detection, the proposed scheme focuses on facial regions and evaluates frame-to-frame similarity to assess authenticity. The effectiveness of the proposed approach is verified using publicly available datasets.