Federated Learning via Augmented Knowledge Distillation for Heterogenous Deep Human Activity Recognition Systems.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Wearable Technology

Background:

• Deep learning models are crucial for Human Activity Recognition (HAR) on wearable devices.
• Training these models requires large, representative datasets.
• Federated Learning (FL) offers a privacy-preserving approach for distributed model training.

Purpose of the Study:

• To address the limitation of standard Federated Learning (FedAvg) in training heterogeneous model architectures.
• To introduce Federated Learning via Augmented Knowledge Distillation (FedAKD) for distributed training of diverse deep learning models.
• To evaluate FedAKD's effectiveness on both tabular and time-series HAR datasets.

Main Methods:

• Proposed Federated Learning via Augmented Knowledge Distillation (FedAKD) for heterogeneous model training.
• Evaluated FedAKD on waist-mounted tabular and wrist-mounted time-series HAR datasets.
• Compared FedAKD against standard FL (FedAvg) and other model-agnostic FL methods.

Main Results:

• FedAKD enables collaborative training of heterogeneous deep learning models with varying capacities.
• Achieved a 200x reduction in communication overhead compared to FL methods communicating gradients/weights.
• Demonstrated performance gains of up to 20% for clients relative to other model-agnostic FL methods.
• Showed increased robustness under statistically heterogeneous scenarios.

Conclusions:

• FedAKD offers a flexible and efficient solution for privacy-preserving distributed training of heterogeneous HAR models.
• The method significantly improves communication efficiency and model performance in FL settings.
• FedAKD is a promising approach for real-world applications of HAR on wearable devices.