ICIP 2024
We’re thrilled to announce that our paper, “On the Exploitation of DCT-Traces in the Generative-AI Domain” has been accepted at the ICIP 2024 conference, which will be held in Abu Dhabi from October 27 to October 30.
In this paper, we analyzed deepfake images generated by GAN and Diffusion Models in the frequency domain, focusing on the statistical distribution of Discrete Cosine Transform (DCT) coefficients. We hypothesized the presence of a unique “discriminative fingerprint” embedded in specific combinations of these coefficients. To identify them, we trained Machine Learning classifiers on various coefficient combinations and employed the Explainable AI (XAI) LIME algorithm to pinpoint intrinsic discriminative combinations. Additionally, we conducted robustness tests by applying JPEG compression to assess the persistence of these traces. Our experimental results demonstrate that generative models leave traces that are both highly discriminative and resilient to JPEG compression.
Git repo will be public as soon as possible.
Arxiv preprint: https://arxiv.org/abs/2402.02209
Arxiv preprint: https://arxiv.org/abs/2402.02209
Authors: Orazio Pontorno, Luca Guarnera, Sebastiano Battiato.