Adversarial domain adaptation for cross-user activity recognition via noise diffusion model

Human Activity Recognition (HAR) is essential for intelligent applications requiring contextual awareness. However, HAR models often face challenges due to data distribution disparities between training and real-world scenarios, particularly across different users. To address this, we propose Diffusion-based Noise-centered Adversarial Learning Domain Adaptation (DNA-DA), a novel framework integrating generative diffusion modeling with adversarial learning for robust cross-user HAR. DNA-DA leverages a tailored network architecture with specialized constraints to align feature distributions across user domains, embedding activity and domain information into noise during diffusion to enhance adaptation. Through adversarial learning, it transforms forward diffusion and reverse denoising into domain alignment phases, enabling robust activity classification. Evaluated on the OPPT, PAMAP2, and DSADS datasets, DNA-DA achieves a 4.0% average accuracy improvement over state-of-the-art methods, with noise-based denoising enhancing data quality. By effectively mitigating distribution mismatches, DNA-DA offers a scalable solution for real-time, user-adaptive HAR systems.