A Secure Hierarchical Federated Learning Using Dirichlet-based Trust Management

Hierarchical Federated Learning (HFL) is a distributed machine learning training system in which a server works with several clients and edge nodes while maintaining data privacy. Distributed machine learning training systems are also known as Federated Learning, but HFL is a type of Federated Learning that utilizes a hierarchical network architecture to address computational issues when dealing with a high number of clients. However, HFL is vulnerable to attacks such as data poisoning, which may jeopardize the entire training process and result in misclassifications. As system defenders, we have to tackle this issue. Using a label-flipping attack, we investigate the effect of data poisoning attacks on HFL training. We propose a trust management-based strategy to mitigate data poisoning attacks, which assesses client trustworthiness using a Dirichlet distribution. We maintain a record of previous activities, allowing the server to enhance its knowledge based on client reliability. We demonstrate the proposed approach's effectiveness through improvements in model performance after removing malicious clients, using the MNIST dataset as a benchmark.