横向联邦学习中PCA差分隐私数据发布算法

为了让不同组织在保护本地敏感数据和降维后发布数据隐私的前提下,联合使用PCA进行降维和数据发布,提出横向联邦PCA差分隐私数据发布算法。引入随机种子联合协商方案,在各站点之间以较少通信代价生成相同随机噪声矩阵。提出本地噪声均分方案,将均分噪声加在本地协方差矩阵上。一方面,保护本地数据隐私;另一方面,减少了噪声添加量,并且达到与中心化差分隐私PCA算法相同的噪声水平。理论分析表明,该算法满足差分隐私,保证了本地数据和发布数据的隐私性,较同类算法噪声添加量降低。实验从隐私性和可用性角度评估该算法,证明该算法与同类算法相比具有更高的可用性。

PCA differential privacy data publishing algorithm in horizontal federated learning

To allow different organizations to jointly use PCA for dimensionality reduction and data release under the premise of protecting the privacy of local sensitive data and the data released after dimensionality reduction, this paper proposed a horizontal federated PCA differential privacy data publishing algorithm. It introduced a random seed joint negotiation scheme to generate the same noise matrix between sites with less communication cost. It proposed a local noise averaging scheme, and added the averaging noise to the local covariance matrix. On the one hand, it protected local data privacy. On the other hand, it reduced the amount of noise added and achieved the same noise level as the centralized differential privacy PCA algorithm. Compared with similar algorithms, the amount of noise added was reduced. The experiment evaluates the algorithm from the perspective of privacy and utility, and proves that the algorithm has higher utility compared with similar algorithms.