Researchers find speech tokenizers have transferable semantic attack vulnerabilities

1. 10:02
   XiaohongshuXiaohongshu · Wang Shuo (SJTU)· Discussion

   Can speech tokenizers also be misled? Transferable semantic attacks(opens the original post in a new tab)

1. 10:33
   arXiv: 其他 AI 安全论文

   Speech tokenizers are fragile: transferable semantic attacks and robust tokenizers

   Researchers found that speech tokenizers are vulnerable at the semantic level: small perturbations can disrupt their semantic encoding, causing severe performance degradation in downstream ASR and large audio language models (LALMs). The paper proposes T-SemAttack, a transferable semantic attack based on an ensemble of semantic encoders, which jointly drifts the representation spaces of multiple semantic encoders to disrupt semantic content while preserving speech perceptual quality. Analysis of layer-wise representation drift and cross-model gradient alignment shows that perturbations can cause large shifts in deep semantic representations, and that the input gradient directions of different surrogate encoders are consistent, suggesting that shared adversarial-sensitive directions in the input space may be a key factor in the transferability of cross-model attacks. The paper also proposes a robust speech tokenizer, ROSETok, combining noise training with robust semantic distillation to improve reconstruction fidelity and the robustness of downstream SpeechLLMs. Experiments cover multiple datasets, 9 open-source and black-box ASR systems, and 4 LALMs; code and demos are at https://wxzyd123.github.io/t-semattack。

Current attention 37·Peak in the comparable period 37 (Nov 15 11:00)·Change over the comparable past 24 hours +3%