Summary
Detecting fraud on networks that change over time is hard because labeled examples often come late or not at all and natural changes make old detection methods unreliable. The authors found a simple way to spot fake or harmful nodes by checking if their features fit well with their neighbors' features, calling this idea feature-context consistency. They built this into a mathematical model that can detect fraud without needing labeled examples and keeps working well even as the network changes naturally. Their tests on real-world data, like Ethereum scams and Reddit banned users, showed their method performs better than many other approaches without needing any labeled data.
What this means in practice
- •For blockchain security teams: Detect phishing accounts in cryptocurrency transaction networks without needing labeled fraud examples and maintain accuracy as the network changes.
- •For online community moderators: Identify banned or malicious users in social platforms by spotting inconsistencies in their attributes and neighborhood behavior without manual labeling.
Abstract
Adversarial detection on evolving attributed graphs faces two challenges: fraud labels arrive late or not at all, and natural distributional drift erodes any trained decision boundary. Modern detectors can reach strong detection rates when labels are available and the distribution is stationary; sustaining that quality without labels and under drift is the open problem. We identify feature-context consistency -- the alignment between a node's attributes and its aggregated neighborhood attributes -- as a single structural signal that addresses both. We formalize it as the Context Boost (CB) score within a Restricted Boltzmann Machine framework (CB-RBM) and prove two complementary worst-case guarantees: a strict separation between the expected CB of legitimate and adversarial nodes, with gap growing linearly in model capacity, whenever the adversary's context stays far enough from legitimate neighborhood context; and a Wasserstein-Lipschitz bound on how far the CB distribution moves under natural drift. We validate on XBlock Ethereum phishing (~153K nodes) and Reddit banned-user detection (~11K nodes), with two further datasets in the appendix spanning diverse domains. With zero labels, CB-RBM achieves AUROC >= 0.96 on four adversarial attacks and keeps FPR close to its calibrated 5% target under drift -- outperforming supervised GNNs (hundreds of labels), unsupervised graph-anomaly baselines, and a standard RBM whose FPR collapses to 89-100%. What typically requires separate mechanisms -- label-free operation and drift stability -- here follows from a single structural regularity of the data. Code: https://github.com/serval-uni-lu/CB-RBM.