Universal Adversarial Attack for Sybil Detection System

Given the widespread adoption of graph-based classification methods across academia and industry, ensuring the robustness of these methods against adversarial attacks has emerged as a significant challenge in graph learning. Currently, adversarial attacks typically involve altering the graph structure through the addition of spurious nodes or manipulating node connections to deceive classifiers. However, few studies have conducted research on finding universal nodes to improve the universality of adversarial attacks.This paper introduces the inaugural endeavor into universal adversarial attacks within the domain of Sybil detection systems. The attacker’s strategy involves identifying a small set of universal nodes and manipulating their connections to deceive the Sybil detection system. We frame this attack as a graph-based optimization problem, necessitating the adjustment of edge weights to achieve the attacker’s objectives. Our assessment of these attacks across multiple datasets underscores their efficacy in avoiding graph-based classification methods.