∊-k anonymization and adversarial training of graph neural networks for privacy preservation in social networks

Abstract

With the explosive growth of social networks, privacy preservation as a social good has been one common concern. Graph neural networks (GNNs) have been utilized by social network service providers to improve business service. However, traditional anonymization techniques of social networks cannot satisfy the desired privacy preservation of node attribute and graph structure and introduce information disturbance from the anonymization, leading to the performance degradation of GNNs in social network analysis. To protect sensitive user data and persist GNNs’ performance in social network analysis, we propose a two-stage privacy-preserving method of graph neural networks in the social network domain. During the first stage, we design a novel ∊-k anonymization method that can achieve ∊-local differential privacy (∊-LDP) and k-degree anonymity by incorporating the classical LDP and k-degree anonymization (k-DA) while retaining as much network community information as possible. At the second stage, we develop an adversarial training mechanism for GNNs to resist the disturbance from ∊-k anonymization and retain as much task performance as possible on anonymous social network data. Comprehensive experiments on several real-world social network datasets demonstrate the effectiveness of the proposed method for privacy-preserving node classification, link prediction, and graph clustering in social networks. The proposed method represents an interesting and important combination of classical anonymous technologies and recent GNNs and can preserve user privacy while providing business service.