Publication

Abstract : This paper addresses the issues of blockchain platforms, especially Ethereum, smart contract vulnerabilities have resulted in significant monetary losses. The primary objective of this research is to identify critical smart contract vulnerabilities, such as reentrancy, overflow, and access control problems, by utilizing Graph Neural Networks (GNNs). To evaluate the degree to which two GNN-based models work at reducing practical financial losses, two of them are created using various training approaches. The first model treats each vulnerability class equally in an effort to maximize accuracy. The second model takes a more cost-effective tack by giving training greater weight to major weaknesses. We investigate the trade-offs between accuracy and financial risk reduction using a dataset of 2,217 contracts with imbalanced class distribution, replicating real-world settings. Financial loss estimates are derived from industry-reported incidents, and confusion matrices are employed to demonstrate the detection patterns. The results show that even though the first model's overall accuracy is higher at 96%, it endures more financial loss as a result of missing vital vulnerabilities. With a 94% accuracy rate, on the other hand, the cost-sensitive model effectively lowers high-impact false negatives, resulting in a less overall financial loss. Given that even a slight drop in accuracy might lead to more successful financial risk mitigation, our findings emphasize the significance of prioritizing vulnerability identification based on possible real-world repercussions. The groundwork for incorporating cost-sensitive models into smart contract auditing tools to improve blockchain security is provided by this research.