Abstract: 

Digital image forensics aims to determine the authenticity of digital images. When looking for traces of manipulation, established forensic methods rely on artifacts from conventional lossy compression, e.g., JPEG. While these traces are often imperceptible, currently proposed neural compression methods introduce artifacts at a semantic level. Reconstructions of image details may appear plausible and of high visual quality but can differ semantically from the original input. My research focuses on the reliability of existing image forensics methods when applied to new (i.e., untested) implementations of compression algorithms or entirely new compression paradigms.

In this talk, I will present recent work on detecting specific artifacts of trellis quantization. The detector is intended to prevent misclassifications that can occur when established forensic methods are applied to images compressed with recent JPEG implementations that use this optimization feature. I will also show examples of semantic changes to image details caused by neural compression and propose a taxonomy that should facilitate work on mitigating the problem.