MixedNUTS: Training-Free Accuracy-Robustness Balance via Nonlinearly Mixed Classifiers

TL;DR

MixedNUTS is a training-free ensemble method that enhances accuracy and robustness of classifiers by nonlinearly mixing their logits, effectively balancing the two without retraining models.

Contribution

It introduces a novel training-free approach to improve accuracy-robustness trade-off by nonlinear mixing of logits from robust and standard classifiers.

Findings

Significantly improves CIFAR-100 accuracy by 7.86 points.

Achieves near state-of-the-art robustness with minimal accuracy loss.

Demonstrates effectiveness on CIFAR-10, CIFAR-100, and ImageNet datasets.

Abstract

Adversarial robustness often comes at the cost of degraded accuracy, impeding real-life applications of robust classification models. Training-based solutions for better trade-offs are limited by incompatibilities with already-trained high-performance large models, necessitating the exploration of training-free ensemble approaches. Observing that robust models are more confident in correct predictions than in incorrect ones on clean and adversarial data alike, we speculate amplifying this "benign confidence property" can reconcile accuracy and robustness in an ensemble setting. To achieve so, we propose "MixedNUTS", a training-free method where the output logits of a robust classifier and a standard non-robust classifier are processed by nonlinear transformations with only three parameters, which are optimized through an efficient algorithm. MixedNUTS then converts the transformed logits into probabilities and mixes them as the overall output. On CIFAR-10, CIFAR-100, and ImageNet datasets, experimental results with custom strong adaptive attacks demonstrate MixedNUTS's vastly improved accuracy and near-SOTA robustness -- it boosts CIFAR-100 clean accuracy by 7.86 points, sacrificing merely 0.87 points in robust accuracy.