Bayesian Source Attribution of Synthetic Images

The authenticity of digital images has become increasingly important, particularly for the question whether an image is photographic (``real'') or generated by an AI system. Also the provenance of an AI-generated image may be an important cue in a forensic analysis. The task of deepfake source attribution aims to distinguish K + 1 classes, namely whether an image is real or generated by one out of K image generators. This work proposes a method for source attribution with a Bayesian fusion of binary detectors with zero-shot priors. Priors from three recently proposed zero-shot statistics are compared, namely the LPIPS distance, latent cosine similarity, and coding-cost gaps. Our results show that the accuracy in a 8-class scenario increases about 8.1 % compared to the baseline detector. Furthermore, the false negative rate for a binary detection task of generated images increases by 10.6 %, which may aid applications like fraud detection. Code is available at https://github.com/vaidya-apoorva/bayesian-ai-image-attribution.