Introduction
Credit ratings in the interbank market are fundamental to global financial stability. However, individual financial institutions often have strong incentives to manipulate their perceived creditworthiness to secure better lending terms or regulatory standing. Such adversarial actions, known as "poisoning attacks," can lead to systemic risks, as evidenced by the 2008 financial crisis.
Attackers typically employ two strategies: Feature Attacks, where internal financial statements are falsified (e.g., inflating assets), and Structural Attacks, where banks create fraudulent lending relationships with high-rated institutions to gain "prestige by association." While Graph Neural Networks (GNNs) are powerful tools for modeling interbank relationships, they are highly susceptible to these manipulations. This paper introduces SA-GNN (Selective-aware Graph Neural Network), a robust framework designed to maintain rating accuracy in the face of such data poisoning.
Methodology: The SA-GNN Framework
The SA-GNN architecture addresses vulnerabilities through three primary mechanisms: a Selective Representation Layer, Target Similarity constraints, and Joint Optimization.
Selective Representation Layer
To mitigate feature attacks, SA-GNN employs a stochastic filtering mechanism. Using a Bernoulli distribution, the model randomly "drops" nodes across multiple independent training runs. This process allows the model to identify and prioritize the most consistent sub-graphs, effectively "guessing" and filtering out poisoned nodes that introduce high prediction errors.
Target Similarity (Label & Feature)
The framework leverages the principle of homophily—the idea that similar entities tend to connect. It introduces two penalty terms:
- Label Similarity (LS): Ensures that connected banks should have logically similar ratings.
- Feature Similarity (FS): Ensures that banks with identical ratings share similar financial profiles.
By penalizing "outlier" relationships or features that deviate from these financial logics, the model makes it significantly harder for attackers to successfully fake high-status connections.
Joint Optimization
The final training objective is a weighted combination of classification loss, feature similarity, and label similarity. This joint optimization allows the model to adaptively focus on legitimate data while penalizing suspicious inconsistencies.
Experimental Evaluation
The authors evaluated SA-GNN using a massive, real-world global interbank dataset spanning 2011 to 2020, encompassing over 20,000 banks and 142,720 data points. The dataset includes diverse bank types and critical financial indicators like assets, liabilities, and impaired loan rates.
Robustness Against Attacks
The model was tested against standard GNNs (GCN, GAT) and specialized defense models (RGCN, Pro-GNN) under varying attack rates from 0% to 25%.
As shown in the results, SA-GNN demonstrates superior robustness. At a 10% perturbation rate, the accuracy decline for SA-GNN was significantly lower than that of the baseline GCN. In a 2019 test scenario with a 20% attack rate, SA-GNN maintained an accuracy of 60.07%, whereas the baseline GCN's performance dropped to 49.89%. Ablation studies further confirmed that both the Selective Representation and Target Similarity modules are essential for this high level of defense.
Conclusion
SA-GNN provides a novel and effective defense against adversarial attacks in interbank credit rating systems. By combining stochastic node filtering with financial logic-based similarity constraints, it offers a resilient alternative to traditional GNNs. Future work will focus on the challenges of obtaining perfectly "clean" training data and exploring the generalizability of the selective-aware approach to other graph-based domains like social networks and recommendation systems.
Summary updated