Privacy-preserving Federated Learning for Resisting Reconstruction Attacks
Huichao Wang, Dacang Feng · 2021
With the development of big data, artificial intelligence has widely walked to aspects of life, bringing life scientific and technological convenience, but also hidden dangers, the most important of which is the issue of data privacy. For privacy issues, researchers focus on data protection, while ignoring reconstruction attacks caused by the leakage of parameters such as gradients in machine learning. In fact, the attacker can indeed recover the original training data through model gradient parameters. In order to address the privacy problem, we design a privacy-preserving federated learning based on homomorphic encryption, which can resist reconstruction attacks. We analyze the correctness of the designed system theoretically and prove its accuracy experimentally with public datasets. The experimental results show that our algorithm performs a good effect while ensuring high accuracy and considerable computation time.