Defending against Poisoning Attack in Federated Learning Using Isolated Forest
Siman Huang, Yan Zhi Bai, Zehua Wang, Peng Liu · 2022
Federated learning is an emerging, distributed machine learning methodology in which participants submit parameters after using their own local data for training and the server aggregates model parameters submitted by users to form a federated learning model. By sending the local learning results rather than sending out the raw data to the server, the user cannot see how the global model is generated. This opacity protects user privacy. However, since the raw data is never sent to the server, its quality and integrity cannot be guaranteed. Malicious attackers can inject poisoning attacks into their own data or the uploading parameters. Thus, federated learning is very vulnerable to distributed poisoning attacks, and this causes the model performances to be reduced significantly. Designing a robust federated learning is a challenging task. In this paper, we propose an isolated forest-based federated learning defense model as IFFed to eliminate the poisoned model parameters of attackers. In each iteration of federated learning, we use isolated forests to partition the data space of model parameters and calculate the probability of anomaly for each participant. Further, we use auxiliary data to design dynamic thresholds to exclude attackers and minimize the impact of the defense model on benign users. Experiments show that IFFed can automatically defend against poisoning attacks, and the performance of the defending model can be significantly improved by adjusting the pre-trained auxiliary model to ensure proper training of the global model.