.png){kind=icon}
.png){kind=icon}
Return
MKDD-Vul: A lightweight multi-modal knowledge distillation framework for detecting vulnerabilities in smart contracts
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.130619.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Smart contracts are vital in decentralized applications, yet vulnerability attacks have led to significant financial losses. Deep learning-based vulnerability detection methods rely on one or a few different modalities (e.g., graph representation, sequence representation, image representation) and have achieved promising performance. However, some of these methods have not completely leveraged interaction information between modalities and employed complex neural networks, which makes them difficult to apply in resource-constrained environments. To end this, we propose MKDD-Vul, a lightweight multi-modal knowledge distillation framework for vulnerability detection, which utilizes multi-modal features (i.e., control flow graph representation, grayscale image representation, and opcode sequence representation) of the smart contract to capture diverse vulnerability features. Specifically, MKDD-Vul employs a dynamic attention knowledge distillation mechanism, which dynamically adjusts the weights of each teacher model based on their features, thereby effectively distilling complementary knowledge from all modalities into a lightweight student model, leading to more accurate and efficient vulnerability detection. We conduct experiments on two datasets, and the experimental results demonstrate that MKDD-Vul achieves performance gains in multiple vulnerability types. For example, the F1 values of MKDD-Vul are improved by 1.62 %, 1.98%, and 5.89% over the state-of-the-art methods for timestamp vulnerabilities, reentrancy vulnerabilities, and overflow vulnerabilities, respectively. Furthermore, we use ablation studies to confirm the efficacy of each module. The code and data for this paper are available on GitHub at: https://github.com/1qq1p/MKDD-Vul
Keywords:
Smart contract
Vulnerability detection
Multimodal learning
Knowledge distillation
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W