@Article{cmc.2025.074959,
AUTHOR = {Yiming Yu, Yunfei Guo, Junchen Liu, Yiping Sun, Junliang Du},
TITLE = {Interpretable Smart Contract Vulnerability Detection with LLM-Augmented Hilbert-Schmidt Information Bottleneck},
JOURNAL = {Computers, Materials \& Continua},
VOLUME = {87},
YEAR = {2026},
NUMBER = {2},
PAGES = {--},
URL = {http://www.techscience.com/cmc/v87n2/66600},
ISSN = {1546-2226},
ABSTRACT = {Graph neural networks (GNNs) have shown notable success in identifying security vulnerabilities within Ethereum smart contracts by capturing structural relationships encoded in control- and data-flow graphs. Despite their effectiveness, most GNN-based vulnerability detectors operate as black boxes, making their decisions difficult to interpret and thus less suitable for critical security auditing. The information bottleneck (IB) principle provides a theoretical framework for isolating task-relevant graph components. However, existing IB-based implementations often encounter unstable optimization and limited understanding of code semantics. To address these issues, we introduce ContractGIB, an interpretable graph information bottleneck framework for function-level vulnerability analysis. ContractGIB introduces three main advances. First, ContractGIB introduces an Hilbert–Schmidt Independence Criterion (HSIC) based estimator that provides stable dependence measurement. Second, it incorporates a CodeBERT semantic module to improve node representations. Third, it initializes all nodes with pretrained CodeBERT embeddings, removing the need for hand-crafted features. For each contract function, ContractGIB identifies the most informative nodes forming an instance-specific explanatory subgraph that supports the model’s prediction. Comprehensive experiments on public smart contract datasets, including ESC and VSC, demonstrate that ContractGIB achieves superior performance compared to competitive GNN baselines, while offering clearer, instance-level interpretability.},
DOI = {10.32604/cmc.2025.074959}
}