USENIX Security2025Top-tier venue
SpeechGuard: Recoverable and Customizable Speech Privacy Protection
Jingmiao Zhang, Suyuan Liu, Jiahui Hou, Zhiqiang Wang, Haikuo Yu, Xiang-Yang Li
Abstract
Uploading speech data to cloud servers poses privacy risks, making the protection of both acoustic and content privacy essential. Users often need the cloud to process non-sensitive information while protecting sensitive parts, with the ability to recover original data locally. However, achieving speech privacy protection that supports fine-grained customization and full recoverability remains a significant challenge. Existing methods often rely on irreversible or inflexible techniques, such as uniformly anonymizing the entire speech or replacing sensitive texts, making them inadequate for this purpose. We introduce SPEECHGUARD, a system that enables recoverable and customizable speech privacy protection by applying reversible protection methods and assigning private information to permission groups. We design a multi-parameter warping function with an inverse function for voice conversion to protect acoustic privacy. We also develop a mechanism for automatic or manual detection and encryption of sensitive texts to protect content privacy. By categorizing listeners into permission groups and assigning warping parameters and encryption keys, SPEECHGUARD enables different listeners to recover varying levels of acoustic and content information according to their permissions, ensuring personalized access to speech data. Experiments on three datasets show that SPEECHGUARD outperforms three baseline systems in anonymity, sensitive content confidentiality, and attack resistance. Moreover, it provides recoverable and customizable protection for acoustic and content privacy, allowing for tailored privacy definitions and protection strength.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- Who is Real Bob? Adversarial Attacks on Speaker Recognition Systems cites CommanderSong: A Systematic Approach for Practical Adversarial Voice Recognition
- V-Cloak: Intelligibility-, Naturalness- & Timbre-Preserving Real-Time Voice Anonymization cites Who is Real Bob? Adversarial Attacks on Speaker Recognition Systems
Builds on5
- wav2vec 2.0: A Framework for Self-Supervised Learning of Speech RepresentationsAlexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, Michael AuliNeurIPS 2020 · 9,451 citations
- CommanderSong: A Systematic Approach for Practical Adversarial Voice RecognitionXuejing Yuan, Yuxuan Chen, Yue Zhao, Yunhui Long et al.USENIX Security 2018 · 389 citations
- Who is Real Bob? Adversarial Attacks on Speaker Recognition SystemsGuangke Chen, Sen Chen, Lingling Fan, Xiaoning Du et al.S&P 2021 · 239 citations
- Preech: A System for Privacy-Preserving Speech TranscriptionShimaa Ahmed, Amrita Roy Chowdhury, Kassem Fawaz, Parmesh RamanathanUSENIX Security 2020
- V-Cloak: Intelligibility-, Naturalness- & Timbre-Preserving Real-Time Voice AnonymizationJiangyi Deng, Fei Teng, Yanjiao Chen, Xiaofu Chen et al.USENIX Security 2023
Related papers
- IO-RAE: Information-Obfuscation Reversible Adversarial Example for Audio Privacy ProtectionJiajie Zhu, Xia Du, Xiaoyuan Liu, Ji-Zhe Zhou et al.AAAI 2026
- VoiceCloak: Adversarial Example Enabled Voice De-Identification with Balanced Privacy and UtilityMeng Chen, Li Lu, Junhao Wang, Jiadi Yu et al.UbiComp 2023 · 21 citations
- WaveGuard: Understanding and Mitigating Audio Adversarial ExamplesShehzeen Hussain, Paarth Neekhara, Shlomo Dubnov, Julian J. McAuley et al.USENIX Security 2021 · 89 citations
- Overo: Sharing Private Audio RecordingsJaemin Lim, Kiyeon Kim, Hyunwoo Yu, Suk-Bok LeeCCS 2022 · 2 citations
- For Human Ears Only: Preventing Automated Monitoring on Voice DataIrtaza Shahid, Nirupam RoyUSENIX Security 2025
