Graph Neural Networks (GNNs) have powerful representation capabilities for graph data, achieving excellent performance across various fields. Considering the scarcity of labels in real-world scenarios, graph self-supervised learning (GSSL) has gained increasing attention due to its ability to train without relying on labels. However, recent studies have revealed that GNNs are vulnerable to stealthy backdoor attacks in GSSL scenarios, enabling the encoder to learn backdoor features simply by injecting triggers. Existing graph backdoor defense methods mainly focus on supervised settings and cannot be directly transferred to self-supervised scenarios due to the lack of label guidance. To bridge this gap, we proposeGDetox, the first backdoor defense approach against backdoored encoders in GSSL.GDetoxaims to eliminate backdoor logic in encoders while maintaining the encoder's original performance. Specifically,GDetoxcan purify the graph backdoor encoder based on the self-supervised distillation approach without relying on label information. Further, we introduce an adversarial contrastive learning that augments node representations without relying on labels to enhance teacher model performance, thereby improving distilled encoder performance. We evaluate the defense performance ofGDetoxon four node classifications and four graph classification datasets by comparing with four state-of-the-art (SOTA) defense methods against seven latest backdoor attack methods on GSSL. Extensive experiments demonstrate thatGDetoxfar outperforms the SOTA defense methods, reducing the attack success rate to 4% with negligible degradation in encoder performance (within 2%) in both node-level and graph-level tasks.

Graph Neural Networks (GNNs) are highly susceptible to numerous adversarial attacks, among which the backdoor attack is one of the toughest to deal with due to the fact that it can lead to misclassification of the model. Similar to Deep Neural Networks (DNNs), backdoor attacks in GNNs work by an attacker changing a portion of the graph data with a hidden trigger and modifying their labels to target labels, which induces the model to learn the trigger feature during its training phase. Although recent defense techniques have emerged, approaches based on explainability and data isolation often fail to detect malicious samples with covert triggers, while discrepancy learning methods tend to degrade performance by removing useful features. To overcome these limitations, we propose a novel backdoor defense method, namedGraphCleanse, on GNNs that can effectively eliminate the possible backdoor features during the training process. Specifically,GraphCleansecan easily break the strong correlation between backdoor features and target labels based on graph contrastive training. To further improve the model accuracy, we present a mutual information maximization method to learn the important feature information in the labeled credible samples and unlabeled suspicious samples by clustering the features obtained from the graph contrastive encoder. Compared with the potential solutions, such as randomized smoothing,GraphCleanseeffectively avoids the negative influence of backdoored samples while maintaining a high model performance. Extensive experimental evaluations on four benchmark datasets demonstrate thatGraphCleansecan reduce the attack success rate to 10% with less performance degradation (within 7%).

Phishing scams are one of Ethereum's most representative security risks that can defraud many transactions in a short period and severely threaten network security. Existing deep learning-based phishing scam detection methods mainly rely on constructing static transaction graphs which are assumed to be accessible before model training. However, static methods that have a high false positive rate to detect newly generated phishing scams by adding this newly generated data to existing algorithms for execution, due to new accounts and transactions constantly appearing in the real-world Ethereum network. Therefore, this article, for the first time, proposes a novel evolve-based phishing scams detection method (named GrabPhisher) that extracts temporal features of accounts and captures information about the dynamic topology of the graph as it evolves. Specifically, GrabPhisher can build the evolutionary pattern of accounts trading on Ethereum as a diffusion network graph in continuous time. It can continue to capture new transaction features based on existing transactions, which facilitates the identification of phishing accounts. Additionally, we implement GrabPhisher on the real-world Ethereum phishing scams datasets. Extensive experimental results demonstrate that GrabPhisher can effectively extract dynamic temporal features and outperform state-of-the-art methods (95% Recall, and 88% F1-score).