Feature-Context Consistency: Unsupervised Adversarial Detection with Drift Stability on Attributed Graphs

Key points

• 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.1
• 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.1
• 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.1
• 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%.1

Sources (1)

• [1]Feature-Context Consistency: Unsupervised Adversarial Detection with Drift Stability on Attributed Graphs
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 12:30 PM

  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.

Extractive summary: sentences quoted from the sources.