On the Exploitation of DCT-Traces in the Generative-AI Domain

Generative AI models leave detectable fingerprints: Both GAN and Diffusion Model-based deepfakes contain unique statistical traces in their Discrete Cosine Transform (DCT) coefficients that distinguish them from authentic images, suggesting that synthetic content is not perfectly indistinguishable from real content.

Not all technical features matter equally for detection: The research found that specific combinations of DCT coefficients are more discriminative than others, meaning deepfake detection systems can be made more efficient by focusing on the most informative technical signatures rather than analyzing all available data.

Detected traces survive common image manipulation: The fingerprints identified in this study remained persistent even after JPEG compression was applied, indicating that deepfakes retain detectable markers even when subjected to common image processing techniques that bad actors might use to evade detection.

Explainable AI helps identify what matters: By using the LIME algorithm (an Explainable AI tool), researchers could identify which specific coefficient combinations were most important for detection, making the detection process more transparent and interpretable rather than a "black box."

Open-source tools advance the field: The authors made their code and dataset publicly available, enabling other researchers and security professionals to build upon these findings and develop more robust defenses against deepfakes.