FedECPA: An Efficient Countermeasure Against Scaling-Based Model Poisoning Attacks in Blockchain-Based Federated Learning

Artificial intelligence (AI) and machine learning (ML) have become integral to various applications, leveraging vast amounts of heterogeneous, globally distributed Internet of Things (IoT) data to identify patterns and build accurate ML models for predictive tasks. Federated learning (FL) is a distributed ML technique developed to learn from such distributed data while ensuring privacy. Nevertheless, traditional FL requires a central server for aggregation, which can be a central point of failure and raises trust issues. Blockchain-based federated learning (BFL) has emerged as an FL extension that provides guaranteed decentralization alongside other security assurances. However, due to the inherent openness of blockchain, BFL comes with several vulnerabilities that remain unexplored in literature, e.g., a higher possibility of model poisoning attacks. This paper investigates how scaling-based model poisoning attacks are made easier in BFL systems and their effects on model performance. Subsequently, it proposes FedECPA-an extension of FedAvg aggregation algorithm with Efficient Countermeasure against scaling-based model Poisoning Attacks in BFL. FedECPA filters out clients with outlier weights and protects the model against these attacks. Several experiments are conducted with different attack scenarios and settings. We further compared our results to a frequently used defense mechanism, Multikrum. Results show the effectiveness of our defense mechanism in protecting BFL from these attacks. On the MNIST dataset, it maintains an overall accuracy of 98% and 89% and outperforms our baseline with 4% and 38% in both IID and non-IID settings, respectively. Similar results were achieved with the CIFAR-10 dataset.