DETER: Detecting Edited Regions for Deterring Generative Manipulations

With the recent booming of Generative AI, concerns regarding the potential malicious usage of deep generative models and their generated fake data have reached their zenith. Existing deep fake datasets are mainly designed for detecting classic editing operations like GANs-based face swapping, presenting two major defects when deploying the detection system trained on them in real-life scenarios. On the one hand, prior construction techniques lag behind the current revolutionary advances in generative modeling that move towards versatile and fine-grained manipulations such as arbitrary image region inpainting via Diffusion Models. On the other hand, we note that they tend to introduce spurious correlations in the benchmark evaluation suites that result in a high false alarm rate with current detection methods. To this end, we introduce the DETER, featuring an up-to-date, high-quality, and large-scale image dataset for DETEcting edited image Regions and deterring more generic generative manipulations. DETER includes 300,000 images manipulated by four state-of-the-art generators and integrates three editing operations in diverse granularities (i.e., face swapping, inpainting, and attribute editing) in realistic visual scenarios. Notably, we conduct extensive experiments and break-down analysis using our rich annotations and improved benchmark protocols, revealing the future direction and challenge for supporting reliable regional fake detection models.