Towards Universal Robust Federated Learning via Meta Stackelberg Game

Henger Li, Tao Li, Yunian Pan, Tianyi Xu, Quanyan Zhu, Zizhan Zheng

Primary Area: societal considerations including fairness, safety, privacy

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Federated Learning, Game Theory, Reinforcement Learning, Robust Machine Learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

Recent studies have revealed that federated learning (FL) systems are susceptible to a range of security threats. Although various defense mechanisms have been proposed, they are typically non-adaptive and tailored to specific types of attacks, leaving them insufficient in the face of unknown/uncertain or adaptive attacks. In this work, we formulate adversarial federated learning as a Bayesian Stackelberg Markov game (BSMG) to tackle adaptive attacks of uncertain types. We further develop an efficient meta-learning approach to solve the game, which provides a robust and adaptive FL defense.
Theoretically, we show that our algorithm provably converges to the first-order $\varepsilon$-equilibrium point in $O(\varepsilon^{-2})$ gradient iterations with $O(\varepsilon^{-4})$ samples per iteration. Empirical results show that our meta-Stackelberg framework obtains superb performance against strong model poisoning and backdoor attacks with uncertain types.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4076