.png){kind=icon}
.png){kind=icon}
Return
Efficient Path-Driven Vulnerability Detection for Complex Smart Contracts via Graph Compression
.png){kind=icon}
PRE
AI
DOI：10.1007/s10922-026-10052-7.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Smart contracts, as core components of blockchain systems, encapsulate the automated execution logic of decentralized applications. Their vulnerabilities, however, can lead to severe security breaches and financial losses, making efficient automated vulnerability detection critically important. Modern smart contracts feature increasingly complex business logic, including extensive function calls and intricate inter-module interactions, which cause the number of potential execution paths to grow exponentially. This path explosion imposes substantial computational challenges on existing detection methods, limiting both efficiency and accuracy. In this work, we propose a graph compression-based path sequence modeling framework that addresses these challenges by simplifying contract graph structures while preserving essential semantic and control dependencies. Our approach extracts candidate subgraphs according to predefined rules and compresses them using a pretrained heterogeneous graph neural network. Sequential features along execution paths are captured with an LSTM encoder, and structural embeddings are fused with path semantics to produce informative representations for downstream analysis. We further introduce a dedicated model, the Boundary-Aware Path Transformer, specifically designed for path-level representation learning. To mitigate the information loss inherent in graph compression, we design a Boundary Alignment Module. This module utilizes an attention mechanism to strictly align the exit features of predecessor nodes with the entry features of successor nodes, thereby ensuring semantic consistency and flow continuity along the compressed execution paths. Experimental results show that SmartPath consistently surpasses state-of-the-art approaches, achieving a 32.75% average F1-score improvement on the synthetic cross-contract vulnerability dataset and a further 10.21% gain on real-world complex contract detection.
Keywords:
Smart contract
Ethereum
Blockchain
Vulnerability detection
Deep learning
Journal
IF:
3.9
Papers:
1.0K
Citations:
1.3K
Organization
View more organizations