The impact of neural image compression on expert practices in digital image forensics

DegreeMaster
StatusActive
Supervisor(s)Nora Hofer, MA

Description

Neural image compression replaces conventional operators of lossy image compression with learnable elements. Current approaches achieve unprecedented compression rates by optimizing perceptual image quality rather than pixel-level fidelity. As a result, neural compression can introduce visually plausible yet semantically incorrect image content, a phenomenon known as miscompression. While such artifacts may go unnoticed by lay users, they pose new challenges for digital image forensics, where experts rely on subtle visual cues.

The objective of this thesis is to investigate how neural image compression and miscompressions affect expert analyses in digital image forensic practice. Through a qualitative interview study with forensic experts, the student will examine how miscompressions might influence analytical strategies, the interpretation of image content, and confidence in the outcome of forensic analyses. The interviews will be supported with demonstration material designed to explain neural compression principles to forensic investigators. The study seeks to identify and assess potential risks posed by an uptake of neural compression.

References

  • Hofer, N. and Böhme, R. A Taxonomy of Miscompressions: Preparing Image Forensics for Neural Compression. In IEEE International Workshop on Information Forensics and Security (WIFS). IEEE, Rome, Italy, 2024, pp. 1–6. [PDF] [Publisher]
  • Mentzer, F., Toderici, G.D., Tschannen, M., and Agustsson, E. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33, (2020), 11913–11924.