FedCNO

Smart contract vulnerability detection is a critical task in ensuring the security of blockchain systems. However, challenges such as non-public source code and label noise, especially systematic noise, significantly hinder the effectiveness of detection models.  This paper introduces FedCNO (Federated Contract Noise Optimizer), an innovative federated learning framework designed to address these issues. FedCNO uniquely integrates local and global label correction mechanisms to improve label accuracy while maintaining data privacy. Additionally, it introduces a consistency score method to dynamically adjust loss weights based on sample reliability, ensuring robust learning even in high-noise environments. Through these mechanisms, FedCNO effectively mitigates the impact of both random and systematic noise, demonstrating superior robustness and adaptability compared to existing approaches. Under 30\% random noise, FedCNO achieved an F1 score of 87.86\% on the CBGRU model. Under 30\% systematic noise, it achieved an F1 score of 72.59\%. In contrast, other baseline methods failed to improve the performance of detection models under systematic noise and even caused significant performance degradation. Experimental results demonstrate that FedCNO exhibits exceptional robustness in handling both random and systematic label noise, significantly outperforming state-of-the-art federated learning approaches.  Beyond noise correction, the framework promotes collaborative model optimization across distributed datasets without compromising sensitive information. Future work will further explore its applicability to diverse types of vulnerabilities and other blockchain security challenges.

智能合约漏洞检测是保障区块链系统安全的核心任务。然而，非公开源代码与标签噪声（尤其是系统性噪声）等问题，严重制约了检测模型的效能。本文提出FedCNO（Federated Contract Noise Optimizer，联邦合约噪声优化器）——一种专为解决上述问题设计的创新性联邦学习框架。该框架创新性地融合了本地与全局标签修正机制，在保障数据隐私的同时提升标签准确率。此外，其还提出了一种一致性评分方法，可基于样本可靠性动态调整损失权重，确保在高噪声环境下仍能实现稳健学习。通过上述机制，FedCNO可有效缓解随机噪声与系统性噪声带来的影响，相较现有方法展现出更优异的鲁棒性与适配性。在30%随机噪声场景下，FedCNO在CBGRU模型上取得了87.86%的F1值；在30%系统性噪声场景下，其F1值达到72.59%。与之相比，其他基线方法在系统性噪声环境下不仅未能提升检测模型的性能，甚至会导致性能显著下降。实验结果表明，FedCNO在处理随机与系统性标签噪声方面展现出极强的鲁棒性，大幅优于当前最先进的联邦学习方法。除噪声修正外，该框架还可在不泄露敏感信息的前提下，推动分布式数据集间的协同模型优化。未来工作将进一步探索其在各类漏洞及其他区块链安全问题中的应用潜力。