FedCon: A Model Consistency-Based Mechanism to Safeguard Federated Learning in Vulnerable and Heterogeneous IoT Environments

Federated Learning (FL) offers a privacy-preserving and cost-efficient paradigm to collaboratively train a global model without exposing local data of clients in IoT. However, since client data remain decentralized, it makes the learning vulnerable to malicious behaviors, such as data poisoning attacks, through which, updates can be manipulated to corrupt the global model and damage the overall IoT system. To address such a security threat, gradient similarity-based methods have been proposed to remedy the impact of data poisoning in securing FL, mainly under independent and identically distributed (IID) scenarios. Since heterogeneous Non-IID data are widely distributed in IoT systems, and direct aggregation on local models is more native for FL to update the global model, it becomes challenging for current solutions to distinguish whether models are learned from clients with poisoned or imbalanced data samples. Therefore, this paper proposes FedCon, a model consistency-based mechanism to secure the FL process. It leverages the consistency of local models to identify benign models for global model aggregation. Experimental results demonstrate that FedCon can outperform state-of-the-art methods in defending against data poisoning attacks under both IID and Non-IID scenarios to not only boost model accuracy but also reduce attack success rate.