On the Adversarial Robustness of Mixture of Experts
This addresses the problem of enhancing adversarial robustness in neural networks for AI safety and reliability, offering a practical approach through MoEs, though it is incremental as it builds on existing robustness and MoE research.
The paper investigates whether sparse Mixture of Experts (MoE) models, which scale parameters without increasing computational cost, improve adversarial robustness compared to dense models. Theoretically, MoEs can have smaller Lipschitz constants under certain conditions, and empirically, they show greater robustness on ImageNet with adversarial attacks.
Adversarial robustness is a key desirable property of neural networks. It has been empirically shown to be affected by their sizes, with larger networks being typically more robust. Recently, Bubeck and Sellke proved a lower bound on the Lipschitz constant of functions that fit the training data in terms of their number of parameters. This raises an interesting open question, do -- and can -- functions with more parameters, but not necessarily more computational cost, have better robustness? We study this question for sparse Mixture of Expert models (MoEs), that make it possible to scale up the model size for a roughly constant computational cost. We theoretically show that under certain conditions on the routing and the structure of the data, MoEs can have significantly smaller Lipschitz constants than their dense counterparts. The robustness of MoEs can suffer when the highest weighted experts for an input implement sufficiently different functions. We next empirically evaluate the robustness of MoEs on ImageNet using adversarial attacks and show they are indeed more robust than dense models with the same computational cost. We make key observations showing the robustness of MoEs to the choice of experts, highlighting the redundancy of experts in models trained in practice.