The study of human activity recognition has numerous benefits for important fields in day-to-day activities. Along with the availability of mobile devices, the benefits of human activity recognition applications can be widespread to numerous clients. ...
The study of human activity recognition has numerous benefits for important fields in day-to-day activities. Along with the availability of mobile devices, the benefits of human activity recognition applications can be widespread to numerous clients. However, realizing HAR in privacy-sensitive, large-scale environments remains challenging due to data privacy concerns and the inherent heterogeneity of distributed data sources. Federated Learning (FL) provides a promising solution by enabling collaborative model training across multiple clients without sharing raw data. Nevertheless, the implementation of federated learning in HAR faces additional challenges related to real-world heterogeneity, such as statistical, participation, and data quality (noisy labels) heterogeneity that negatively impact the performance of human activity recognition.
This thesis addresses these key limitations through three approaches. First, to mitigate statistical heterogeneity, a federated framework combining contrastive learning and adaptive control variates is proposed. Contrastive learning aligns local and global representations for faster convergence, while adaptive control variates regulate local updates to stabilize training.
Second, to address client participation heterogeneity, an FL framework inspired by the mixture-of-experts paradigm integrates a global model with personalized local models. The global model captures generalizable patterns, whereas the local model learns user-specific behaviors. For predictions, a fusion strategy combines the strengths of both models, leading to improved classification performance.
Finally, a new framework with a two-stage training scenario is proposed to perform FL training with clients that carry noisy labels. The first stage identifies separating the client into clean and noisy client groups. The second stage consists of separate training for both groups. In each round, a scheduled model threshold, along with a noisy partition weight, helps to carefully integrate the corrected noisy label into the overall FL training. Additionally, a new model structure that utilizes noisy labels as input is proposed to learn robust invariant features along with the feature-label relation to enhance the prediction of the true label.
Comprehensive experiments on multiple benchmark HAR datasets validate the effectiveness of the proposed methods under various real-world constraints.