PriMonitor: An adaptive tuning privacy-preserving approach for multimodal emotion detection

Abstract

The proliferation of edge computing and the Internet of Vehicles (IoV) has significantly bolstered the popularity of deep learning-based driver assistance applications. This has paved the way for the integration of multimodal emotion detection systems, which effectively enhance driving safety and are increasingly prevalent in our daily lives. However, the utilization of in-vehicle cameras and microphones has raised concerns regarding the extensive collection of driver privacy data. Applying privacy-preserving techniques to a single modality alone proves insufficient in preventing privacy re-identification when correlated with other modalities. In this paper, we introduce PriMonitor, an adaptive tuning privacy-preserving approach for multimodal emotion detection. PriMonitor tackles these challenges by proposing a generalized random response-based differential privacy method that not only enhances the speed and data availability of text privacy protection but also ensures privacy preservation across multiple modalities. To determine suitable weight assignments within a given privacy budget, we introduce pre-aggregator and iterative mechanisms. Our PriMonitor effectively mitigates privacy re-identification due to modal correlation while maintaining a high level of accuracy in multimodal models. Experimental results validate the efficiency and competitiveness of our approach.