Private (Stochastic) Non-Convex Optimization Revisited: Second-Order Stationary Points and Excess Risks

We consider the problem of minimizing a non-convex objective while preserving the privacy of the examples in the training data. Building upon the previous variance-reduced algorithm SpiderBoost, we introduce a new framework that utilizes two different kinds of gradient oracles. The first kind of ora...

Olles dieđut

Bibliográfalaš dieđut

Váldodahkkit:	Ganesh, Arun, Liu, Daogao, Oh, Sewoong, Thakurta, Abhradeep
Materiálatiipa:	Artihkal
Almmustuhtton:	2023
Fáttát:	Machine Learning Cryptography and Security Optimization and Control
Liŋkkat:	https://arxiv.org/abs/2302.09699 http://arxiv.org/abs/2302.09699
Fáddágilkorat:	Lasit fáddágilkoriid Eai fáddágilkorat, Lasit vuosttaš fáddágilkora!