Theoretical modeling study demonstrates optimal gradient privacy protection in machine learning training, highlighting viable defenses against curious learners and eavesdroppers.
Privacy protection of shared gradients in machine learning model training is crucial, yet the coexistence of honest-curious learners and malicious eavesdroppers in the real world is overlooked in research. A linear strategy-based privacy protection method for three-party games with eavesdroppers is proposed to address privacy threats in machine learning model training. Firstly, a three-party game framework is designed based on machine learning model training, with the data owner as the leader and the learner and eavesdropper as followers, leading to a privacy protection optimisation problem. Secondly, two eavesdropping scenarios are considered: limited and complete encoded information. By solving the equilibrium solution of the optimisation problem, the optimal strategy pair is obtained, and a strategic privacy protection information transmission mechanism is proposed. Finally, simulation experiments illustrate the theoretical results, validating the effectiveness of the algorithms proposed in this paper.
No takes yet. Share an insight, caveat, or question.
Jiang et al. (2025) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: