Publikationen

Autor	Steinebach, Martin; Yannikos, York
Datum	2025
Art	Conference Paper
Abstrakt	Robust image hashing algorithms are designed to produce similar or equal hash values for image variants despite benign transformations like resizing or compression. However, modern generative techniques like image inpainting undermine this robustness by introducing semantically meaningful content changes that are imperceptible at the pixel level. This paper demonstrates how such inpainting attacks can exploit perceptual hash functions, causing significant semantic alterations to go undetected or, conversely, generating false positives that overload detection systems. We evaluate several widely used hashing schemes including Blockhash, rHash, ISCC, and PhotoDNA and show that they struggle to detect inpainting-based manipulations while remaining overly sensitive to trivial changes like cropping. These findings reveal a critical trade-off in hash function design: the balance between robustness to benign edits and sensitivity to malicious content alteration. We discuss the implications for systems relying on hash-based content detection, such as those targeting child sexual abuse material (CSAM), and highlight the potential for denial-of-service and strike inflation attacks. Our results underscore the urgent need for semantically aware image hashing solutions capable of withstanding adversarial generative modifications.
Konferenz	International Conference on Availability, Reliability and Security 2025
Url	https://publica.fraunhofer.de/handle/publica/490711