Robust Aggregation Function in Federated Learning

Maintaining user data privacy is a crucial challenge for machine learning techniques. Federated learning is a solution that enables machine learning models to be trained using data residing on different devices without centralizing the data. This training method offers several advantages: Firstly, federated learning helps preserve user privacy by storing data on separate devices rather than transferring it to a central location for training. Secondly, training machine learning models on a diverse set of devices through federated learning improves their robustness, ensuring optimal performance in a wide range of real-world scenarios. Finally, federated learning can promote scalability by enabling simultaneous training on a vast number of devices. So, it can increase the scale of training and enable the development of more sophisticated models. The aggregation function is one of the main steps in federated learning and is used on the server side to aggregate local models sent from the client side. The most widely used aggregation function is Krum, which, despite the research done to improve its robustness, is still vulnerable to adversarial samples. In this paper, a method is proposed to improve the robustness of the Krum aggregation function. The results confirm that the proposed method is more robust against adversarial samples than the original version of the Krum aggregation function.