Skip to content

BadGCL: Persistent Backdoor Attacks Against Graph Continual Learning Systems

BadGCL: Towards Persistent Backdoor Attack on Graph Continual Learning Systems

arXiv: 其他 AI 安全论文· arXiv:2611.19656· Dong Li, Yufei Han, Xintao Wu, Xujiang Zhao, Wei Cheng, Chen ZhaoBaylor University; INRIA; University of Arkansas; Meta; NEC Laboratories AmericaIndexed by this site · 2 days ago· Originally publishedAttacks· Its place in the scan record →

Translation in progress — untranslated text is omitted.

Quick read

Attacks

Based on the paper PDF (AI-generated); refer to the original

AI summaryDong Li and other researchers propose BadGCL, a model-agnostic persistent backdoor attack framework against graph continual learning (GCL) systems. The paper points out that GCL's knowledge retention mechanism also retains backdoors injected early when poisoned updates are compatible with clean task signals, and that structural triggers embedded in the message-passing paths of graph neural networks (GNNs) cause stable representation shifts across tasks, amplifying this risk. Methodologically, under a strict early-intervention setting, BadGCL uses a gradient mutual information objective to learn topology-aware structural triggers, couples poisoned gradients with the clean task optimization signal, and uses a propagation consistency objective to stabilize the effect of the triggers through message passing; theoretical analysis shows that gradient compatibility and propagation stability jointly support the long-term persistence of the backdoor under benign continual updates. Experiments on task-IL and class-IL protocols, four GCL benchmarks and multiple victim systems show that BadGCL maintains a persistent attack effect with limited degradation in clean utility, significantly outperforms graph backdoor and continual learning backdoor baselines, and remains robust to representative defenses.

Questions
In graph continual learning (GCL) systems, existing graph backdoor attacks fail as tasks are updated, giving the impression that continual updates naturally weaken backdoors. The authors argue that GCL's knowledge retention mechanism, when exploited, may instead retain early-injected backdoors over the long term.
Methods
Proposes BadGCL: intervenes only in the first task, first trains a clean reference model, then uses gradient mutual information (implemented with InfoNCE) to couple the poisoning gradient with the clean task signal, and uses a propagation consistency objective to stabilize the trigger's effect in message passing, supplemented by structural regularization, and finally exits after a one-time poisoning fine-tuning.
Experiments and results
On 4 GCL benchmarks, under the task-IL and class-IL protocols, and on 4 victim systems, BadGCL's final ASR is mostly higher than the baselines, e.g., CoraFull/PDGNNs reaches 99.46% under task-IL and 98.74% under class-IL (Table 1, 2); among 5 defenses, A2GBD is the strongest, reducing ASR to 89.42% under class-IL (Table 4a).
Limitations and future work
The paper does not specifically discuss limitations. Experimental coverage: there are 4 victim systems in total—PDGNNs, CaT-CGL, SSRM, and TPP; defense evaluation is performed only on CoraFull+PDGNNs; the task-order and task-revisiting experiments are each conducted on a single dataset, Reddit and OGB-Arxiv respectively; CBACL and BTB were excluded because their threat models do not match.
Read originalarxiv.org