Frequency-Domain Analysis of Traces for the Detection of AI-based Compression

Abstract

The JPEG algorithm is the most popular compression method on the internet. Its properties have been extensively studied in image forensics for examining image origin and authenticity. However, the JPEG standard will in the near future be extended with AI-based compression. This approach is fundamentally different from the classic JPEG algorithm, and requires an entirely new set of forensics tools.As a first step towards forensic tools for AI compression, we present a first investigation of forensic traces in HiFiC, the current state-of-the-art AI-based compression method. We investigate the frequency space of the compressed images, and identify two types of traces, which likely arise from GAN upsampling and in homogeneous areas. We evaluate the detectability on different patch sizes and unseen postprocessing, and report a detectability of 96.37%. Our empirical results also suggest that further, yet unidentified, compression traces can be expected in the spatial domain.