PUPPY: Perturbation-aware Topological Purification with Proxy Rectification for Test-time Graph Anomaly Detection

Graph anomaly detection (GAD) aims to identify anomalous patterns in graph-structured data, with conventional methods typically relying on the independent and identically distributed (i.i.d.) assumption. In real-world scenarios, previously unseen yet normal patterns may emerge during deployment, introducing normality shifts that degrade detection performance. However, existing test-time adaptation methods remain unsatisfactory due to heterogeneous semantic pattern conflicts and biased predictions under normality shifts. % However, normality patterns learned from training graphs may not remain valid during deployment. In real-world scenarios, previously unseen but normal patterns can emerge in both attributes and graph structure, Towards this end, in this paper, we propose a novel approach named Perturbation-aware Topological Purification with Proxy Rectification () for test-time GAD. The core of our proposed is to jointly suppress structural contamination and correct prediction bias under normality shifts. Given the pretrained GAD detector, on the one hand, we estimate the sensitivity of each node under graph perturbations, which further guides selective edge pruning for test graph structure transformation. On the other hand, we introduce a self-distillation framework that leverages a proxy model trained on high-confidence target nodes as a reliable teacher. In this way, we model the adapting detector’s prediction drift relative to the pretrained source model and debias the student prediction with mutual adaptation distillation. Extensive experiments across diverse test-time graph anomaly detection benchmarks demonstrate the effectiveness of our proposed against the baseline method. Our code is available at https://anonymous.4open.science/r/PUPPY-code-94F2/https://anonymous.4open.science/r/PUPPY-code-94F2/.