Robust Federated Learning via Relevant Model Slicing

Federated Learning (FL) is a machine learning paradigm in which many nodes collaborate to learn shared global models using local datasets. Nowadays, Federated Learning is widely used in the explosion of data. However, in addition to the shortcomings of the models themselves in machine learning, Federated Learning is also more vulnerable to backdoor attacks due to its own open nature, i.e., it is public during the training process and generally involves a large number of participants, any of whom may have malicious behaviors. Reducing or even countering these attacks is very challenging, as existing methods increase the aggregation cost or drastically reduce the task accuracy. In this paper, we propose FedLRMS (Federated Learning via Relevant Model Slicing). Specifically, it computes model weights related to the respective client's data based on the effective neuron information obtained by analyzing the global model usage, and utilizes only the weights that are relevant to the client's data for fine-tuning the personalized Federated Learning model, while choosing to re-initialize and re-train the model weights for irrelevant ones. Through FedLRMS, we significantly mitigate the impact of backdoor attacks on the training process. In order to verify the performance of FedLRMS, the experiments utilize two datasets, MNIST, CIFAR-10, and use two models, ResNet18 and ResNet50. Two kinds of non-independent homogeneous distribution scenarios, data imbalance and category imbalance, as well as independent homogeneous distribution scenarios, are considered. In addition, the number of backdoor clients is increased from 40% to 60% in 10% increments, in order to realize the scenarios with a large number of backdoor clients, and the FedLRMS algorithm is also tested in the scenarios with fewer backdoor clients than normal clients, and the experimental results show that FedLRMS is able to resist the backdoor attacks in different scenarios completely, and the success rate of backdoor attacks is only about 0.2% (0% in some scenarios), and at the same time, it obtains a better accuracy of the main task, in contrast, several existing classical algorithms fail completely, and the success rate of the backdoor attack is only about 2%. In contrast, several existing classical algorithms fail completely, mostly making the success rate of backdoor attacks 100%, which demonstrate the superiority of the FedLRMS algorithm.