Trimmed Averaging for Efficient Federated Learning in the Internet of Things

The adoption of Federated Learning (FL) for decen-tralised model training has enabled workload distribution across multiple client devices. However, extreme values and outliers in the model updates from distributed devices often challenge this approach. This study introduces the Adaptive Trim strategy (AdaTrim), which leverages a trimmed mean aggregation to enhance robustness against outliers and adversarial attacks. The strategy is tested on two data sets to show its impact, one from the Industrial Internet of Things (IIoT) and the Internet of Medical Things (IoMT). A comprehensive evaluation framework also provides a holistic assessment of the model's performance. Our results demonstrate the efficacy of AdaTrim in maintaining high model accuracy while ensuring robustness and reliability in diverse FL scenarios. Furthermore, we discuss the adoption of AdaTrim, highlighting its efficiency in improving security for the Internet of Things (IoT) while preserving data privacy.