Research on Deep Learning Based on Decentralized Differential Privacy Protection
Quan Zhou, Yongchang Lao, Yongliang Yin, Wei Cao · 2024
Most existing decentralized learning algorithms focus on improving the convergence speed and communication efficiency of models, without paying too much attention to the issue of participant privacy leakage. Differential privacy is currently one of the mainstream privacy protection technologies in the field of deep learning, widely used to protect the data privacy and security of participants. This article proposes a decentralized deep learning algorithm (PriDCNN) for differential privacy protection. The PriDCNN algorithm is based on a decentralized network topology, which avoids the problem of communication traffic congestion in the central node. Nodes cooperate with each other to train deep learning models, resulting in improved communication efficiency. For users with smaller local datasets, this approach can improve the performance of the local model. At the same time, the algorithm protects the training data of participants from being leaked by adding noise. The process of adding noise to PriDCNN is only used as a preprocessing method for the model and is not limited by the number of iteration rounds, which is beneficial for model optimization.