Improving the Accuracy-Robustness Trade-Off of Classifiers via Adaptive Smoothing
Authors:
- 14 August 2024
- journal article
- research article
- Published by Society for Industrial & Applied Mathematics (SIAM) in SIAM Journal on Mathematics of Data Science
- Vol. 6 (3) , 788-814
Abstract
While prior research has proposed a plethora of methods that build neural classifiers robust against adversarial robustness, practitioners are still reluctant to adopt them due to their unacceptably severe clean accuracy penalties. Real-world services based on neural networks are thus still unsafe. This paper significantly alleviates the accuracy-robustness trade-off by mixing the output probabilities of a standard classifier and a robust classifier, where the standard network is optimized for clean accuracy and is not robust in general. We show that the robust base classifier’s confidence difference for correct and incorrect examples is the key to this improvement. In addition to providing empirical evidence, we theoretically certify the robustness of the mixed classifier under realistic assumptions. We then adapt an adversarial input detector into a mixing network that adaptively adjusts the mixture of the two base models, further reducing the accuracy penalty of achieving robustness. The proposed flexible mixture-of-experts framework, termed “adaptive smoothing,” works in conjunction with existing or even future methods that improve clean accuracy, robustness, or adversary detection. We use strong attack methods, including AutoAttack and adaptive attacks, to evaluate our models’ robustness. On the CIFAR-100 dataset, we achieve an [Formula: see text] clean accuracy while maintaining a [Formula: see text] [Formula: see text]-AutoAttacked ([Formula: see text]) accuracy, becoming the second most robust method on the RobustBench benchmark as of submission, while improving the clean accuracy by 10 percentage points over all listed models. Code implementation is available at https://github.com/Bai-YT/AdaptiveSmoothing .Keywords
All Related Versions
Funding Information
- Office of Naval Research
- National Science Foundation
- UC Noyce Initiative
- C3.ai Digital Transformation Institute
This publication has 32 references indexed in Scilit:
- Efficient Global Optimization of Two-Layer ReLU Networks: Quadratic-Time Algorithms and Adversarial TrainingSIAM Journal on Mathematics of Data Science, 2023
- Data-Driven Certification of Neural Networks With Random Input NoiseIEEE Transactions on Control of Network Systems, 2022
- Revisiting model’s uncertainty and confidences for adversarial example detectionApplied Intelligence, 2022
- A Differentiable Language Model Adversarial Attack on Text ClassifiersIEEE Access, 2022
- Adversarial example detection for DNN models: a review and experimental comparisonArtificial Intelligence Review, 2022
- An adversarial attack detection method in deep neural networks based on re-attacking approachMultimedia Tools and Applications, 2021
- Square Attack: A Query-Efficient Black-Box Adversarial Attack via Random Search2020
- Big Transfer (BiT): General Visual Representation Learning2020
- Towards Robust Neural Networks via Random Self-ensemble2018
- Adversarial image detection in deep neural networksMultimedia Tools and Applications, 2018