Privacy-preserving global user profile construction through federated learning
Zheng Huo, Teng Wang, Yilin Fan, Ping He · International Journal of Computational Science and Engineering · 2023
User profiles are derived from big data left on the internet through machine learning algorithms. However, threats of data privacy leakages restrict access to the data in centralised machine learning. Federated learning (FL) can avoid privacy leakage during the data collection phase. Herein, we propose an algorithm for constructing a privacy-preserving global user profile (PPGUP) through FL in a vertical data-segmentation scenario. Participants train local clusters on their data using the CLIQUE algorithm, and carefully encrypt cluster parameters using Paillier encryption to protect parameters from an untrusted aggregator. The aggregator then makes intersections over the cluster parameters without decryption, to construct a GUP. The experiment results show that precision of PPGUP reaches 80% when is set to 1.5, which is improved by 50% compared with DP-UserPro. The runtime exhibits a linear growth with the growth of the dataset size and the increase in the number of participants.