Out-of-distribution AI-generated image detection with zero-phase deconvolution for residual whitening
Current AI-generated image detection methods struggle in out-of-distribution testing conditions because detectors are usually not robust when evaluated on generators not included in the training set. This work explores a way to mitigate the forensic bias between sources of images -whether real or generated -by removing generator-specific fingerprints. This new training process can be approximated by pre-processing involving a deconvolution operation designed to remove correlated noise patterns inducing loss of generalization. Indeed, we show that whitening the spatial distribution of the image noise residuals can significantly improve the performance of detectors on images coming from generators unseen during training. In particular, an end-to-end CNN model trained on whitened image residuals improves detection by 12.6% on average compared to a training involving original images.