CLEP: A Novel Contrastive Learning Method for Evolutionary Reentrancy Vulnerability Detection

Authors: Jie Chen, Liangmin Wang, Huijuan Zhu, Victor S. Sheng

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results show that CLEP not only outperforms state-of-the-art baselines in version-specific scenarios but also shows promising performance in cross-version evolution scenarios.
Researcher Affiliation Academia 1School of Cyber Science and Engineering, Southeast University 2School of Computer Science and Communication Engineering, Jiangsu University 3Department of Computer Science, Texas Tech University EMAIL, EMAIL, EMAIL
Pseudocode Yes Algorithm 1: Version-Aware Contrastive Sampler
Open Source Code No The paper does not provide any statement regarding the release of source code, nor does it include a link to a code repository or mention code in supplementary materials.
Open Datasets Yes Datasets. 1) Smart Bug Wild (Durieux et al. 2020) is a widely recognized smart contract dataset. [...] 2) Re Tran Study (Zheng et al. 2023) is a recently cross-version smart contract dataset, which contains 139, 424 contracts.
Dataset Splits Yes Following Peculiar (Wu et al. 2021) and Re Vul DL (Zhang et al. 2022), we split the datasets into training (20%), validation (10%), and testing sets (70%).
Hardware Specification Yes All experiments were conducted with two Intel Xeon 6148 CPUs, two 3090Ti GPUs, and 512GB RAM.
Software Dependencies No The paper does not explicitly state specific software dependencies with version numbers, such as programming languages, libraries, or frameworks used in the implementation.
Experiment Setup Yes The learning rate l tuned amongst {1e 6, 1e 5, 2e 5, 5e 5}, the epoch ep is searched in {4, 6, 8, 10, 15}, code length c in {256, 512}, the batch size bs in {24, 32, 64, 128} and data flow length d in {64, 128}. The performance of all neural network models, unless specifically stated, are reported on the following settings: for VSFT, we set l to 1e 5, ep to 5, bs to 24, c to 256, and d to 64; for CVFT, l was set to 1e 6, ep to 10, bs to 32, c to 256, and d to 64; and for SFT, l, ep, and bs were set to 1e 5, 10, and 32 respectively in the refinement phase, and to 2e 5, 4, and 48 in the fine-tuning phase.