Trustworthy Federated Learning Against Malicious Attacks in Web 3.0
Yuan Zheng, Youliang Tian, Zhou Zhou, Ta Li, Shuai Wang, Jinbo Xiong
Abstract
In the era of Web 3.0, federated learning has emerged as a crucial technical method in resolving conflicts between data security and open sharing. However, federated learning is susceptible to various malicious behaviors, including inference attacks, poisoning attacks, and free-riding attacks. These adversarial activities can lead to privacy breaches, unavailability of global models, and unfair training processes. To tackle these challenges, we propose a trustworthy federated learning scheme (TWFL) that can resist the above malicious attacks. Specifically, we firstly propose a novel adaptive method based on two-trapdoor homomorphic encryption to encrypt gradients uploaded by users, thereby resisting inference attacks. Secondly, we design confidence calculation and contribution calculation mechanisms to resist poisoning attacks and free-riding attacks. Finally, we prove the security of our scheme through formal security analysis, and demonstrate through experiments conducted on MNIST and FASHIONMNIST datasets that TWFL achieves a higher model accuracy of 2%–3% compared to traditional methods such as Median and Trim-mean. In summary, TWFL can not only resist a variety of attacks but also ensure improved accuracy, which is enough to prove that it is a trustworthy solution suitable for Web 3.0 privacy protection scenarios.