Privacy-preserving federated graph neural network against poisoning attack

Graph neural network (GNNs) has gradually moved from theory to application, however less attention has been paid to training for privacy preserving. Due to the particularity of the graph structure, the small disturbance of the graph will also reduce its performance. In order to resist poisoning attacks, this paper proposes a privacy defense strategy based on homomorphic encryption (HE). Specifically, we adopt HE to encrypt local embedding and generate global embedding under ciphertext in order to achieve the confidentiality of node embedding. Secondly, by calculating the cosine similarity between node features in ciphertext. Then the backpropagation process is divided into two parts, which are executed by the user and the server respectively to achieve the privacy of the intermediate gradient. During the whole process, the client’s private data and weights are always invisible to the server. Finally, the theoretical and experimental results show that the proposed protocol has a accuracy error of 1.2%–3.3% compared with the GNN model under plaintext data. Meanwhile, the accuracy of the model with the defense framework could be improved by 22%–27% compared to those models without the defense mechanisms under attack.