Abstract: Most vulnerability detection methods for smart contracts rely on a single modality or large models. The former suffers from insufficient extraction of key features, resulting in low detection accuracy, while the latter has a large parameter scale, which is not conducive to deployment in resourceconstrained environments. To address these issues, this paper proposes dualmodal conditional feature gating, a lightweight dynamic feature fusion method for smart contract vulnerability detection, which adaptively generates fusion weights via a lightweight gating network and constructs a lightweight deploymentoriented detection framework. Experiments on 17,554 reentrancy samples from SmartBugs Wild achieve an accuracy of 76.87%, an F1score of 0.788, and an AUC of 0.870, outperforming static fusion, concatenation fusion, and singlemodal baselines. Compared with reproduced MVul-L and CMRL, the proposed method maintains comparable performance while reducing the parameter count by about 45% and lowering inference latency.

Key words: smart contract, vulnerability detection, dynamic feature fusion, lightweight deployment

摘要： 智能合约漏洞检测方法大多依赖单一模态或大模型，单一模态中关键特征提取不充分导致检测准确率较低，而大模型中参数规模较大不利于资源受限环境下的实际部署。为此，本文提出一种面向轻量化部署的智能合约漏洞双模态动态特征融合检测方法，通过轻量级门控网络自适应生成融合权重，构建面向轻量化部署的智能合约漏洞检测框架。实验表明，测试集准确率达到76.87%，F1-Score达到0.788，AUC达到0.870。与静态加权融合、拼接融合及单模态方法相比均取得更优性能；与多模态方法MVul-L、CMRL等的复现对比表明，本文方法在保持相近检测性能的同时融合与分类模块的可训练参数量降低约45%、推理延迟更低。

关键词: 智能合约, 漏洞检测, 动态特征融合, 轻量化部署