Membership Inference Attacks Against Self-supervised Speech Models

2021/11/09 by Wei-Cheng Tseng, Tseng, Wei-Cheng, Wei-Tsung Kao +3 · 5 citations
Computer Science · Engineering · #Audio and Speech Processing (eess.AS) #Cryptography and Security (cs.CR) #FOS: Computer and information sciences #FOS: Electrical engineering #Machine Learning (cs.LG) #Sound (cs.SD) #cs.CR #cs.LG #cs.SD #eess.AS #electronic engineering #information engineering

paper · pdf · doi:10.48550/arxiv.2111.05113

Accepted to Interspeech 2022. Code will be available in the future

arxiv created 2022/08/15 · arxiv updated 2022/08/16

Abstract

Recently, adapting the idea of self-supervised learning (SSL) on continuous speech has started gaining attention. SSL models pre-trained on a huge amount of unlabeled audio can generate general-purpose representations that benefit a wide variety of speech processing tasks. Despite their ubiquitous deployment, however, the potential privacy risks of these models have not been well investigated. In this paper, we present the first privacy analysis on several SSL speech models using Membership Inference Attacks (MIA) under black-box access. The experiment results show that these pre-trained models are vulnerable to MIA and prone to membership information leakage with high Area Under the Curve (AUC) in both utterance-level and speaker-level. Furthermore, we also conduct several ablation studies to understand the factors that contribute to the success of MIA.

Cited by

• Assessing the Effectiveness of Membership Inference on Generative Music
• Protecting Bystander Privacy via Selective Hearing in Audio LLMs
• Personal Attribute Leakage in Federated Speech Models
• The Man Behind the Sound: Demystifying Audio Private Attribute Profiling via Multimodal Large Language Model Agents
• Adversarial Attacks in Multimodal Systems: A Practitioner's Survey

Related

• SoK: Reducing the Vulnerability of Fine-tuned Language Models to Membership Inference Attacks
• Reducing audio membership inference attack accuracy to chance: 4 defenses
• Quantifying Privacy Risks of Masked Language Models Using Membership Inference Attacks
• Personal Attribute Leakage in Federated Speech Models
• SEF-MK: Speaker-Embedding-Free Voice Anonymization through Multi-k-means Quantization