同态加密
计算机科学
差别隐私
加密
情绪分析
差速器(机械装置)
理论计算机科学
人工智能
计算机安全
数据挖掘
信息隐私
密码学
隐私保护
情报检索
信息隐藏
作者
Hung Le,Shogo Okada,Van‐Nam Huynh
出处
期刊:Neurocomputing
[Elsevier BV]
日期:2026-09-17
卷期号:707: 135161-135161
标识
DOI:10.1016/j.neucom.2026.135161
摘要
Multimodal Sentiment Analysis (MSA) processes facial expressions, vocal patterns, and linguistic choices, all of which constitute sensitive biometric data vulnerable to exploitation by advanced deepfake technologies. Consequently, any system performing this analysis must implement robust measures to protect user privacy. Developing such a system, however, has been difficult due to two main challenges: increased computation and reduced utility (e.g., model accuracy or usefulness). Multimodal learning involves huge matrix calculations, which are not practical for Homomorphic Encryption (HE), while the interactions between the modalities are so subtle that adding Differential Privacy (DP) noise to the input would destroy performance. This work proposes a framework to balance the triangular trade-off between privacy, computation, and utility. Our key observation is that different modalities require different levels of security. We therefore propose a strict privacy protection for highly identifiable modalities and a more relaxed protection for the remaining modality. In particular, visual and audio features are encrypted with homomorphic encryption, and text features are perturbed with differential privacy before being sent to the online model. The model processes the received input without decrypting the sensitive data or being fully aware of its own predictions. While prior work in privacy-preserving MSA focused on DP-based approaches, ours is the first to introduce the use of HE in a computationally acceptable way. Additionally, our approach allows users to control the level of privacy they are willing to give up in exchange for more accurate predictions. Data and code will be released upon publication.
科研通智能强力驱动
Strongly Powered by AbleSci AI