Mitigate Data Poisoning Attack by Partially Federated Learning

Abstract

An efficient machine learning model for malware detection requires a large dataset to train. Yet it is not easy to collect such a large dataset without violating or leaving vulnerable to potential violation various aspects of data privacy. Our work proposes a federated learning framework that permits multiple parties to collaborate on learning behavioral graphs for malware detection. Our proposed graph classification framework allows the participating parties to freely decide their preferred classifier model without acknowledging their preferences to the others involved. This mitigates the chance of any data poisoning attacks. In our experiments, our classification model using the partially federated learning achieved the F1-score of 0.97, close to the performance of the centralized data training models. Moreover, the impact of the label flipping attack against our model is less than 0.02.