Privacy-preserving federated learning (I)

Abstract

Botnet attack is a critical problem in IoT devices. However, current botnet detection technology like the autoencoder model requires centralized training on a large amount of data collected from IoT devices from different home networks, which is not practical because the sensitive information in these data may cause privacy leakage. Individual Learning within the home network also can diminish the performance of the autoencoder. Federated Learning that allows the home networks to learn a shared model with data kept locally is a solution. In this project, we implement the Federated Learning approach on autoencoder for botnet detection and compare Federated Learning's performance with Centralized Training and Individual Learning in terms of prediction accuracy. The Federated Training approach shows a noticeable improvement in prediction accuracy over Individual Learning and a similar performance compared with Centralized Learning. In addition, to further reduce the chance that the model being attacked, the Federated Learning approach is further transformed to a secure aggregation platform. The prediction result of Federated Learning with and without secure aggregation shows that the secure aggregation platform doesn’t affect the performance of Federated Learning.