Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Adversarial Attack on Large Scale Graph

About

Recent studies have shown that graph neural networks (GNNs) are vulnerable against perturbations due to lack of robustness and can therefore be easily fooled. Currently, most works on attacking GNNs are mainly using gradient information to guide the attack and achieve outstanding performance. However, the high complexity of time and space makes them unmanageable for large scale graphs and becomes the major bottleneck that prevents the practical usage. We argue that the main reason is that they have to use the whole graph for attacks, resulting in the increasing time and space complexity as the data scale grows. In this work, we propose an efficient Simplified Gradient-based Attack (SGA) method to bridge this gap. SGA can cause the GNNs to misclassify specific target nodes through a multi-stage attack framework, which needs only a much smaller subgraph. In addition, we present a practical metric named Degree Assortativity Change (DAC) to measure the impacts of adversarial attacks on graph data. We evaluate our attack method on four real-world graph networks by attacking several commonly used GNNs. The experimental results demonstrate that SGA can achieve significant time and memory efficiency improvements while maintaining competitive attack performance compared to state-of-art attack techniques. Codes are available via: https://github.com/EdisonLeeeee/SGAttack.

Jintang Li, Tao Xie, Liang Chen, Fenfang Xie, Xiangnan He, Zibin Zheng• 2020

Related benchmarks

TaskDatasetResultRank
Node ClassificationSQUIRREL vanilla (test)
Miss-classification Rate13.6
322
Node ClassificationCora
Miss-classification Rate (Δ=1)21.47
96
Node ClassificationPubmed
Miss-classification Rate (Δ=1)22.4
96
Node ClassificationCora
Miss-classification Rate20.93
70
Node ClassificationPubmed
Miss-classification Rate18.27
70
Node ClassificationOGB-ARXIV (test)
Misclassification Rate36.67
60
Node ClassificationCHAMELEON vanilla (test)
Miss-classification Rate (Δ=1)27.6
50
Node ClassificationCiteseer
Miss-classification Rate (Δ=1)19.33
40
Node ClassificationChameleon
Misclassification Rate2.13
28
Node ClassificationCora (test)
Error Rate (Δ=1)8.4
28
Showing 10 of 17 rows

Other info

Follow for update