Federated model with contrastive learning and adaptive control variates for human activity recognition

Abstract

Recent attention to privacy issues demands a communication-safe method for training human activity recognition (HAR) models on client activity data. Federated learning (FL) has become a compelling technique to facilitate model training between the server and clients while preserving data privacy. However, classical FL methods often assume independent and identically distributed (IID) data among clients. This assumption does not hold true in practical scenarios. Human activity in real-world scenarios varies, resulting in skewness where identical activities are executed uniquely across clients. This leads to local model objectives drifting away from the global model objective, thereby impeding overall convergence. To address this challenge, we propose FedCoad, a novel federated model leveraging contrastive learning with adaptive control variates to handle the skewness among HAR clients. Model contrastive learning minimizes the gap in representation between global and local models to help global model convergence. During local model updates, the adaptive control variates penalize the local model updates with respect to the model weight and the rate of change from the control variates update. Our experiments show that FedCoad outperforms state-of-the-art FL algorithms on HAR benchmark datasets.

1. Introduction

Sensor-based human activity recognition (HAR) using wearable devices is critical in user-centered applications like healthcare (Wu et al., 2022), smart environments (Bianchi et al., 2019), and fall detection (Mrozek et al., 2020). Federated learning (FL) (McMahan et al., 2017) offers a decentralized and privacy-preserving solution for training models across devices distributed among clients. The FL method works well when the client data are independent and identically distributed (IID) as most of the clients have similar quantities, labels, and features of data. However, humans perform various activities and even have unique patterns. Therefore, real-world HAR data are inherently non-IID since they exhibit some skewness that can degrade model performance.

There are three types of HAR skewness in real-world conditions (Presotto et al., 2022): feature distribution skew, where two clients perform the same activity with different patterns (e.g., walking patterns differ between younger and older individuals); quantity distribution skew, where clients have imbalanced labeled data volumes; label distribution skew, where two clients have different sets of labels. For example, an athlete has more activity labels related to sports than an office worker does. These conditions lead to divergent local objectives during training and hinder global model optimization (Karimireddy et al., 2020; Li X et al., 2020).

Existing methods have sought to address these non-IID challenges. Some methods focus on the model aggregation stage to handle non-IID issues, such as implementing momentum for model aggregation normalization (Hsu et al., 2019) or clustering clients with similar data distributions for grouped updates (Presotto et al., 2022; Guo JL et al., 2024). However, normalization during model aggregation may fail, as the local updates are already skewed and clustering operations require significant computational resources, especially as the number of clients grows. Other methods focus on the local training stage, such as FedProx, which applies L2-norm regularization to constrain local updates (Li T et al., 2020), and SCAFFOLD, which uses control variates to reduce update variance (Karimireddy et al., 2020). While effective in some cases, these methods risk local model overfitting due to the limited diversity of client data in HAR.