Omnipotent Adversarial Training in the Wild

Adversarial training is an important topic in robust deep learning, but the community lacks attention to its practical usage. In this paper, we aim to resolve a real-world challenge, i.e., training a model on an imbalanced and noisy dataset to achieve high clean accuracy and adversarial robustness, with our proposed Omnipotent Adversarial Training (OAT) strategy. OAT consists of two innovative methodologies to address the imperfection in the training set. We first introduce an oracle into the adversarial training process to help the model learn a correct data-label conditional distribution. This carefully-designed oracle can provide correct label annotations for adversarial training. We further propose logits adjustment adversarial training to overcome the data imbalance issue, which can help the model learn a Bayes-optimal distribution. Our comprehensive evaluation results show that OAT outperforms other baselines by more than 20% clean accuracy improvement and 10% robust accuracy improvement under complex combinations of data imbalance and label noise scenarios. The code can be found in https://github.com/GuanlinLee/OAT.

Paper

References (51)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer TYeL5/10 · confidence 4/52023-10-31

Summary

This paper tackles the challenges associated with noisy and imbalanced training data in a scenario of adversarial training. The proposed method is composed of an oracle training process (resampling, refurbishment & split, and contrastive learning) and adversarial training (distribution estimation and logit adjustment). Experimental results using CIFAR-10 and 100 shows much better clean accuracy (+20% points) and robust accuracy (+10% points) than the baseline methods.

Strengths

+ This paper is the first trial to achieve better clean and robust accuracy for adversarial training under the noisy and imbalanced conditions. + Multiple techniques are efficiently combined for better adversarial training and the key components in this work are all reasonable. + The effectiveness of the proposed method is clearly shown with CIFAR-10 and 100 by significantly outperforming the baseline methods in terms of both clean accuracy and robust accuracy.

Weaknesses

1. I am afraid the technical novelty is relatively weak. - Data-resampling (over-sampling, sub-sampling, SMOTE, etc) can be found in many papers. - The label refurbishment and dataset split using k-NN seems essentially the same as in the following paper. If not, the authors should clarify the novelty more clearly. [a] Dara Bahri, Heinrich Jiang, and Maya Gupta. Deep k-nn for noisy labels, Proceedings of Machine Learning Research, 2020. [b] Chen Feng, Georgios Tzimiropoulos, Ioannis Patras. SSR: An Efficient and Robust Framework for Learning with Unknown Label Noise, BMVC, 2022. -- Contrastive self-supervised learning is also well-known. -- Label distribution estimation, i.e., label reassignment, is commonly used as in [b]. -- Logit adjustment is simply taken from [Menon, ICLR21] 2. The experimental validation is weak - The experiments are conducted only with CIFAR-10 and 100. I admit these two datasets are standard, but the authors might want to show more results with larger datasets to give greater impact to the community. - The baselines are all simple adversarial training methods. As the authors might be aware, there are already many methods for noisy labels and imbalanced data. The authors might want to at least apply such methods to the baselines. Otherwise, I am afraid that I cannot say the comparison is fair.

Questions

As pointed out in the weakness part, I have concerns about the technical novelty. Please clarify which part was technically novel. Just as a reference, an adversarial training method for imbalanced datasets has been presented recently. Since this paper was officially published in Oct. 2023, I am NOT requesting the authors to compare their work to this one. Wentao Wang, Harry Shomer, Yuxuan Wan, Yaxin Li, Jiangtao Huang, and Hui Liu. 2023. A Mix-up Strategy to Enhance Adversarial Training with Imbalanced Data. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management (CIKM '23).

Rating

5: marginally below the acceptance threshold

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Reviewer j17w5/10 · confidence 4/52023-11-01

Summary

This paper addressed the challenges of adversarial training from noisy labeled and class imbalanced dataset. A two stage method was proposed to sequentially update oracle model for label correction and adversarial training on imbalanced data. Comparisons to state-of-the-art adversarial training methods demonstrate the effectiveness of the method.

Strengths

Strength: 1. This paper explores a new setting for adversarial training where training data could be contaminated with noise and imbalanced.

Weaknesses

Weakness: 1. The proposed method does not unify adversarial training with learning from noisy and long-tailed data. These two parts are seemly optimized separately. 2. The techniques introduced for learning from noisy and long-tailed data are not new. Re-sampling is a well-known method for learning from imbalanced data. Label refurbishment is also widely adopted in self-training. Contrastive training is also quite matured. The technical contribution is thus very limited. 3. The evaluation of adversarial attack is not strong enough. PGD is a relatively weak attack method by now. Strong methods, e.g. AutoAttack, should be evaluated additionally. 4. The ablation study is weak. It is hard to see which components are most effective from the presentation in Tab.1.

Questions

It is necessary to elaborate more on the contributions and novelties. Evaluations against stronger adversarial attacks are necessary. A clear ablation study is encourage to better analyse effectiveness of proposed method.

Rating

5: marginally below the acceptance threshold

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Reviewer zhDe3/10 · confidence 4/52023-11-04

Summary

This paper aims to develop an adversarial training for a model on an imbalanced and noisy dataset. The proposed method first introduces an oracle to correct the data labels. To overcome the data imbalance challenge, a dataset re-sampling technique is used. To further improve the label correction process, the self-supervised contrastive learning technique is adopted to train the oracle. Then, the adversarial training is carried out.

Strengths

The problems of label noise, sample imbalance, and adversarial sample are simultaneously considered in training a model.

Weaknesses

Sample labeling noise, sample imbalance, and adversarial sample are problems that have been studied separately, and a number of related methods have been proposed. The proposed method in this draft is a simple combination of existing techniques. The sample label correction, data re-sampling, contrastive learning, and adversarial training included in the proposed method are conventional methods. Overall, not much technological innovation has been seen in this manuscript.

Questions

In addition to combining pre-existing technologies to solve multiple problems, what are the original innovations that stem from this manuscript?

Rating

3: reject, not good enough

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

© 2026 NYSGPT2525 LLC