Preferential Selective-Aware Graph Neural Network for Preventing Attacks in Interbank Credit Rating.

Area of Science:

• Financial modeling
• Network security
• Machine learning

Background:

• Accurate bank credit rating is crucial for financial stability and economic growth.
• Existing rating models are vulnerable to deliberate data manipulation (poisoning attacks).
• Attacks can involve manipulated features or network structures, complicating credit assessment.

Purpose of the Study:

• To propose a novel approach, the preferential selective-aware graph neural network (PSAGNN) model.
• To defend Interbank credit rating systems against both feature and structural nontarget poisoning attacks.
• To enhance the robustness and accuracy of credit rating assessments.

Main Methods:

• Developed a phased optimization approach with biased perturbation.
• Explored Interbank preferences and network scale-free properties for adaptive data prioritization.
• Simulated a clean graph by adaptively prioritizing poisoning training data.
• Implemented a weighted penalty on the opposition function to distinguish attackers.

Main Results:

• The PSAGNN model effectively defends against feature and structural poisoning attacks.
• Demonstrated superior performance compared to state-of-the-art baseline methods.
• Validated through extensive experiments on a newly collected Interbank quarter dataset and case studies.

Conclusions:

• The proposed PSAGNN model significantly improves the resilience of Interbank credit rating systems.
• The approach offers a robust solution for mitigating sophisticated data poisoning attacks.
• This work contributes to a healthier financial environment through more reliable credit assessments.