HBMD-FL: Heterogeneous federated learning algorithm based on blockchain and model distillation

Federated learning is a distributed machine learning framework that allows participants to keep their privacy data locally. Traditional federated learning coordinates participants collaboratively train a powerful global model. However, this process has several problems: it cannot meet the heterogeneous model’s requirements, and it cannot resist poisoning attacks and single-point-of-failure. In order to resolve these issues, we proposed a heterogeneous federated learning algorithm based on blockchain and model distillation. The problem of fully heterogeneous models that are hard to aggregate in the central server can be solved by leveraging model distillation technology. Moreover, blockchain replaces the central server in federated learning to solve the single-point-of-failure problem. The validation algorithm is combined with cross-validation, which helps federated learning to resist poison attacks. The extensive experimental results demonstrate that HBMD-FL can resist poisoning attacks while losing less than 3 $$\%$$ of model accuracy, and the communication consumption significantly outperformed the comparison algorithm. Copyright © 2022 The Author(s), under exclusive license to Springer Nature Switzerland AG.