Certifiably Robust Synthetic Image Detectors

Randomized Smoothing (RS) has been proposed as a method to develop deep learning classifiers with certified robustness, i.e., for which a certain level of robustness can be theoretically guaranteed. In this paper, we explore the application of the RS technique in the context of multimedia forensics, focusing on the prominent task of synthetic image detection. Our experiments, carried out on the task of detection of images generated by StyleGAN2 and Latent Diffusion models, reveal that the input pre-processing, the input size and in particular the network architecture have a noticeable impact on the certification performance. In particular, we achieved the best performance with EfficientNetB4, while we found that the certification achieved by detectors based on general-purpose features, namely CLIP, is poor. We also evaluated the performance of the RS synthetic image detectors against common image post-processing, showing that they exhibit strong robustness against a wide variety of processing, even when the distortion introduced by the processing exceeds the one the detectors can provably withstand.