Revisiting adapters with adversarial training

TL;DR

This paper demonstrates that using adapters with adversarial training in Vision Transformers improves clean accuracy, enables model soups for trade-offs, and adapts well to distribution shifts, all with fewer parameters.

Contribution

It shows that adapters suffice for co-training on clean and adversarial inputs without batch statistic separation, improving accuracy and enabling flexible model combinations.

Findings

Improved top-1 accuracy on ImageNet by +1.12%.

Enabled model soups for clean and adversarial trade-offs.

Achieved +4.00% better accuracy on ImageNet variants.

Abstract

While adversarial training is generally used as a defense mechanism, recent works show that it can also act as a regularizer. By co-training a neural network on clean and adversarial inputs, it is possible to improve classification accuracy on the clean, non-adversarial inputs. We demonstrate that, contrary to previous findings, it is not necessary to separate batch statistics when co-training on clean and adversarial inputs, and that it is sufficient to use adapters with few domain-specific parameters for each type of input. We establish that using the classification token of a Vision Transformer (ViT) as an adapter is enough to match the classification performance of dual normalization layers, while using significantly less additional parameters. First, we improve upon the top-1 accuracy of a non-adversarially trained ViT-B16 model by +1.12% on ImageNet (reaching 83.76% top-1 accuracy). Second, and more importantly, we show that training with adapters enables model soups through linear combinations of the clean and adversarial tokens. These model soups, which we call adversarial model soups, allow us to trade-off between clean and robust accuracy without sacrificing efficiency. Finally, we show that we can easily adapt the resulting models in the face of distribution shifts. Our ViT-B16 obtains top-1 accuracies on ImageNet variants that are on average +4.00% better than those obtained with Masked Autoencoders.