Cybersecurity in Decentralized Machine Learning for Battery Management Systems: Threats, Detection, and Defense

Abstract

Federated Learning (FL) enables distributed devices to collaboratively train machine learning models without sharing raw data, making it suitable for Battery Management Systems (BMSs) that estimate battery State of Health (SOH). However, FL remains vulnerable to attacks such as poisoning attacks in which malicious participants manipulate local training or model updates to influence the learned model. This thesis investigates the security of decentralized battery health prediction systems implemented using the FEDn framework and an Adaptive Iterative Clustered Federated Learning (AICFL) architecture. A controlled experimental environment was developed to evaluate the impact of poisoning attacks on both traditional FL and clustered FL. Three attack categories were implemented and analyzed: model poisoning, stealth-oriented poisoning, and targeted backdoor attacks. To support attack analysis, a server-side suspicious-client risk scoring mechanism was developed to identify anomalous client behavior based on model update characteristics collected during training. Experimental results compare the effectiveness of attacks in FL and AICFL environments and examine how clustering influences attack propagation and model robustness. Experimental results show that poisoning attacks can significantly affect model behavior in both FL and AICFL environments. The impact varies across attack types, while the clustered architecture influences how malicious updates propagate through the federation. The proposed risk-scoring mechanism was able to identify clients exhibiting suspicious update patterns during training. The findings provide insights into the security challenges of clustered federated learning systems for battery management applications and contribute practical methods for analyzing malicious behavior in decentralized AI systems.