ICPR 2024
We are proud to announce that our paper, “DeepFeatureX Net: Deep Features eXtractors based Network for discriminating synthetic from real images“, has just been accepted at the ICPR 2024 conference, which will be held in Kolkata, India, from 1 December to 5 December.
In this work, we address a critical challenge in Digital Forensics: distinguishing between real and AI-generated images, particularly those created by deep learning algorithms known as Deepfakes. Our study introduces an innovative method that improves generalization—the ability to accurately classify images generated by previously unseen architectures.
Our approach involves three specialized blocks, or Base Models, each focusing on extracting discriminative features from images produced by Diffusion Models, GANs, or REAL. By utilizing deliberately unbalanced datasets during training, we enhance the model’s effectiveness. The outputs from these blocks are then combined and processed to determine the image’s origin. The results are promising—our method not only demonstrates strong resilience to JPEG compression and other adversarial attacks but also surpasses current state-of-the-art techniques in generalization tests.
ArXiv preprint: https://arxiv.org/abs/2404.15697
Code, models and dataset are available at https://github.com/opontorno/block-based_deepfake-detection
Authors: Orazio Pontorno, Luca Guarnera, Sebastiano Battiato