A Privacy-First Sustainable HCD-LLM Framework for Mental Health Interaction: Integrating Multimodal Perception With Differential Privacy

Abstract

With the rapid development of artificial intelligence (AI) technology, the application of multimodal large language models (LLMs) in mental health support is becoming increasingly prominent. However, existing systems face challenges, including privacy breaches, insufficient sustainability, and fragmented personalization. This paper proposes a Sustainable Human-Centered Design (HCD) Framework (SHF) that emphasizes multimodal perception and privacy-first mental health interactions. Comparative experiments validated that the proposed model, under a fixed privacy protection (\epsilon =1.0), achieved recognition accuracies of 81.6% and 76.8% on different datasets, with F1 scores of 0.80 and 0.76, respectively. In terms of member inference attack protection, the model achieved an attack success rate of 55.8%. The performance changes under different privacy protections were also analyzed. Ablation experiments confirmed the importance of different components, demonstrating the effectiveness of the design. Statistical reliability was added to all experiments to ensure reproducibility. A good balance between effectiveness and privacy was achieved. The research results provide a practical reference for the mental health field within human-robot interaction (HRI).