Information Leakage and Privacy Protection in Industrial Federated Learning
研究了工业联邦学习中梯度聚合和模型聚合两种场景下的数据隐私泄露问题，发现梯度压缩能有效防御反向攻击，而差分隐私对模型聚合攻击无效。
Nowadays, the application of federated learning in the industry has received a lot of attention and research. It achieves joint modeling across factories by sharing models or gradients and breaks factory-level data silos. However, due to the existence of reverse attacks, attackers can still infer the original data from the shared model parameters or gradients, which poses a serious threat to the privacy of industrial data and the practical deployment of industrial federated learning. Nevertheless, the model vulnerability, evaluation mechanism, and privacy protection methods against reverse attacks have not been thoroughly studied. Therefore, data privacy leakage of industrial federated learning is systematically discussed and thoroughly investigated in two specific scenarios in this article. The results show that the reverse attack against gradient aggregation achieves excellent recovery for all variables of industrial data, and gradient compression can limit this attack well. The inverse attack against model aggregation threatens privacy for some variables with poor robustness, but differential privacy (DP) is ineffective in protecting against such attacks. Consequently, when federated learning cooperation between factories is performed, privacy protection methods of gradient compression are needed if gradient aggregation is used. If model aggregation is used, variables that are less robust to model reversal attacks can be prevented by first finding them through experiments.