ترغب بنشر مسار تعليمي؟ اضغط هنا

Adversarial Training Can Hurt Generalization

224   0   0.0 ( 0 )
 نشر من قبل Sang Michael Xie
 تاريخ النشر 2019
والبحث باللغة English




اسأل ChatGPT حول البحث

While adversarial training can improve robust accuracy (against an adversary), it sometimes hurts standard accuracy (when there is no adversary). Previous work has studied this tradeoff between standard and robust accuracy, but only in the setting where no predictor performs well on both objectives in the infinite data limit. In this paper, we show that even when the optimal predictor with infinite data performs well on both objectives, a tradeoff can still manifest itself with finite data. Furthermore, since our construction is based on a convex learning problem, we rule out optimization concerns, thus laying bare a fundamental tension between robustness and generalization. Finally, we show that robust self-training mostly eliminates this tradeoff by leveraging unlabeled data.



قيم البحث

اقرأ أيضاً

Adversarial training is an approach for increasing models resilience against adversarial perturbations. Such approaches have been demonstrated to result in models with feature representations that generalize better. However, limited works have been d one on adversarial training of models on graph data. In this paper, we raise such a question { does adversarial training improve the generalization of graph representations. We formulate L2 and
Adversarial training (AT) based on minimax optimization is a popular learning style that enhances the models adversarial robustness. Noisy labels (NL) commonly undermine the learning and hurt the models performance. Interestingly, both research direc tions hardly crossover and hit sparks. In this paper, we raise an intriguing question -- Does NL always hurt AT? Firstly, we find that NL injection in inner maximization for generating adversarial data augments natural data implicitly, which benefits ATs generalization. Secondly, we find NL injection in outer minimization for the learning serves as regularization that alleviates robust overfitting, which benefits ATs robustness. To enhance ATs adversarial robustness, we propose NoiLIn that gradually increases underline{Noi}sy underline{L}abels underline{In}jection over the ATs training process. Empirically, NoiLIn answers the previous question negatively -- the adversarial robustness can be indeed enhanced by NL injection. Philosophically, we provide a new perspective of the learning with NL: NL should not always be deemed detrimental, and even in the absence of NL in the training set, we may consider injecting it deliberately.
Adversarially trained models exhibit a large generalization gap: they can interpolate the training set even for large perturbation radii, but at the cost of large test error on clean samples. To investigate this gap, we decompose the test risk into i ts bias and variance components and study their behavior as a function of adversarial training perturbation radii ($varepsilon$). We find that the bias increases monotonically with $varepsilon$ and is the dominant term in the risk. Meanwhile, the variance is unimodal as a function of $varepsilon$, peaking near the interpolation threshold for the training set. This characteristic behavior occurs robustly across different datasets and also for other robust training procedures such as randomized smoothing. It thus provides a test for proposed explanations of the generalization gap. We find that some existing explanations fail this test--for instance, by predicting a monotonically increasing variance curve. This underscores the power of bias-variance decompositions in modern settings-by providing two measurements instead of one, they can rule out more explanations than test accuracy alone. We also show that bias and variance can provide useful guidance for scalably reducing the generalization gap, highlighting pre-training and unlabeled data as promising routes.
Recently, learning a model that generalizes well on out-of-distribution (OOD) data has attracted great attention in the machine learning community. In this paper, after defining OOD generalization via Wasserstein distance, we theoretically show that a model robust to input perturbation generalizes well on OOD data. Inspired by previous findings that adversarial training helps improve input-robustness, we theoretically show that adversarially trained models have converged excess risk on OOD data, and empirically verify it on both image classification and natural language understanding tasks. Besides, in the paradigm of first pre-training and then fine-tuning, we theoretically show that a pre-trained model that is more robust to input perturbation provides a better initialization for generalization on downstream OOD data. Empirically, after fine-tuning, this better-initialized model from adversarial pre-training also has better OOD generalization.
Adversarial training can considerably robustify deep neural networks to resist adversarial attacks. However, some works suggested that adversarial training might comprise the privacy-preserving and generalization abilities. This paper establishes and quantifies the privacy-robustness trade-off and generalization-robustness trade-off in adversarial training from both theoretical and empirical aspects. We first define a notion, {it robustified intensity} to measure the robustness of an adversarial training algorithm. This measure can be approximate empirically by an asymptotically consistent empirical estimator, {it empirical robustified intensity}. Based on the robustified intensity, we prove that (1) adversarial training is $(varepsilon, delta)$-differentially private, where the magnitude of the differential privacy has a positive correlation with the robustified intensity; and (2) the generalization error of adversarial training can be upper bounded by an $mathcal O(sqrt{log N}/N)$ on-average bound and an $mathcal O(1/sqrt{N})$ high-probability bound, both of which have positive correlations with the robustified intensity. Additionally, our generalization bounds do not explicitly rely on the parameter size which would be prohibitively large in deep learning. Systematic experiments on standard datasets, CIFAR-10 and CIFAR-100, are in full agreement with our theories. The source code package is available at url{https://github.com/fshp971/RPG}.

الأسئلة المقترحة

التعليقات
جاري جلب التعليقات جاري جلب التعليقات
سجل دخول لتتمكن من متابعة معايير البحث التي قمت باختيارها
mircosoft-partner

هل ترغب بارسال اشعارات عن اخر التحديثات في شمرا-اكاديميا