Yunjie Ge, Jinlong Jiang, Yuchen Liu, Huayi Duan, Shenyi Zhang, Lingchen Zhao, Chao Qun Shen, Qian Wang · ACM SIGSAC Conference on Computer and Communications Security (CCS) 2026 · 2026
DOI: 10.1145/3830454.3832636
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Modern speaker verification systems (SVSs) have been widely deployed in security-critical applications. We show that these systems exhibit a previously underexplored cluster-based vulnerability. We observe that speaker embeddings produced by SVSs naturally organize into well-separated and dense clusters. Moreover, since SVSs authenticate users by thresholding speaker embedding similarity, the acceptance regions of speakers within the same cluster often overlap substantially. As a result, the overlap region introduces a new attack surface: a single utterance placed in such a region can simultaneously satisfy the verification criteria of multiple distinct enrolled users, enabling one-to-many impersonation. Based on this vulnerability, we propose PolyVoice, a dictionary-based impersonation attack that operates under realistic black-box threat models, where the adversary has no access to the internals of the target SVS or to any enrolled users' voice data. PolyVoice generates a compact set of reusable attack samples, each of which can simultaneously impersonate multiple enrolled users by aligning with cluster-level acceptance regions. To address the absence of enrolled user data, PolyVoice first constructs cluster-aligned attack samples from auxiliary public speech, placing them in shared acceptance regions induced by clustered speaker embeddings. Because embeddings vary across SVSs, samples optimized under a single model may fail to generalize. To address this, PolyVoice applies surrogate-guided optimization over a diverse set of models, biasing them toward cluster-central regions that remain stable across architectures. We evaluate PolyVoice on 4,000 previously unseen speakers across 15 black-box SVSs, including open-source models and commercial services. Using only eight pre-generated samples, it achieves an average attack success rate of 22% at the equal error rate threshold and 13% at a 1% false acceptance rate.
No comments yet — start the discussion below.