GenAIDet

The creation of photorealistic synthetic images or the alteration of existing footage is an emerging societal concern that has led to research in automatic detection of generated content and the release of several benchmarks. While previous research has sought to address this issue, the effectiveness of existing solutions remains limited. Detector models trained on datasets of real and generated images often perform poorly when identifying content from GMs not seen during training. This occurs because these models do not learn what a real image inherently looks like; instead, they focus on specific visual artifacts characteristic of the GMs they encountered during training.  Despite the large volume of generated content, the quality of these images is often inadequate, the reason being twofold: images are generated from scratch using only text conditioning; the generative models used are not the current state-of-the-art. Moreover, resolutions and compression are often different between real and fake images, significantly easing the detection of generated content. In this work, the project proposes a new, challenging benchmark for generated content detection that contains highly realistic images thanks to the use of image conditioning and of very effective and recent models. This benchmark will features three levels of image alterations, from subtle manipulations of real images to full generation, with the same distribution of real images for resolutions and compression levels. To assess generalization, it will provide train/test splits where models are visible only at test time. In addition, the team will investigate the use of energy-based models and contrastive losses for detecting generated content. This method will be evaluated against the proposed benchmark and against publicly available datasets to ensure a fair comparison. The benchmark and model will be publicly released.