Skip to content
Open access

Learn Once, Evade Many: Hash-Center Task Construction for Universal NeuralHash Attacks

2026 · IEEE Signal Processing Letters · Vol 33, pp. 4126-4130 · 0 citations · 21 references

Abstract

Perceptual hash matching is an important client-side mechanism for detecting known harmful content in end-to-end encrypted systems, making the robustness of learned perceptual hashing increasingly critical. Existing NeuralHash attacks are predominantly input-specific and require a new optimization for each image. Universal perturbations avoid this repeated cost, but direct sample-wise optimization over facial identities can suffer from incompatible attack directions. We show that the key issue is how transferable tasks are constructed in NeuralHash's projected code space, rather than how sample gradients are merely aggregated. We therefore construct identity centers, sampled sub-centers, and a global center to define local support and population-level query tasks; a lightweight first-order support–query procedure only realizes this task decomposition. At $\tau =17$, the resulting reusable perturbation achieves a 100% attack success rate (ASR), 10 percentage points above the strongest compared universal baseline, with an LPIPS of 0.257. This work is a full access robustness evaluation: after offline learning, the perturbation is reused on unseen identities without labels, center construction, or per-image gradient optimization.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.