Anomaly Detection in Federated Learning: A Comprehensive Study on Data Poisoning and Energy Consumption Patterns in IoT Devices
Abdelkader Tounsi, Osman Salem, Ahmed Mehaoua · 2024
With the increasing adoption of IoT devices for data capture and machine learning (ML) over distributed multiparty data in various domains, ensuring data privacy and security has become a significant challenge. Existing approaches, such as Federated Learning (FL), are emerging as promising technologies enabling collaborative learning among multiple clients while preserving individual data privacy. FL involves aggregating outputs computed by a group of devices at a central aggregator and running iterative algorithms to train a globally shared model. However, malicious attackers among participating clients can intentionally manipulate training data or the trained model, compromising the system’s accuracy and reliability. In this study, we investigate the effectiveness of various anomaly detection techniques to identify data poisoning attempts within FL frameworks and their impact on local model weights and overall performance. Our approach combines monitoring of weight deviations with energy consumption analysis. By leveraging machine learning algorithms and statistical methods, we aim to enhance the robustness of FL systems against adversarial attacks. Our evaluation demonstrates that the Local Outlier Factor (LOF) model excels in detecting weight deviations, while Adaptive Boosting (AB) is most effective for energy consumption anomalies. These results highlight our approach’s capability to monitor data poisoning and other threats in federated learning environments.