Fundamental limits on adversarial robustness

By Alhussein Fawzi, Omar Fawzi, Pascal Frossard. 2015

58 citations recorded by the source corpus.

Subfields: Safety, Robustness & Interpretability, Statistical Learning Theory

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Safety, Robustness & Interpretability

In Statistical Learning Theory

Citations in the corpus

Accessorize to a Crime: Real and Stealthy Attacks on State-of-the-Art Face Recognition (2016, background): “Later work attributes adversarial examples to the non-ﬂexibility of the classiﬁcation models [9, 13].”
Adversarial Examples: Attacks and Defenses for Deep Learning (2017, background): “Contrary to [60], [132] believed that adversarial examples are due to the “low flexibility” of the classifier for certain tasks.”
Adversarial Manipulation of Deep Representations (2015, methodology): “Recent papers have shown that deep neural networks (DNNs) for image classification can be fooled, often using relatively simple methods to generate so-called adversarial images (Fawzi et al., 2015; Goodfellow et al., 2014; Gu & Rigazio, 2014; Nguyen et al., 2015; Szegedy et al., 2014; Tabacof & Vall”
Threat of Adversarial Attacks on Deep Learning in Computer Vision: A Survey (2018, background): “[117] introduced a framework for studying the instability of classifiers to adversarial perturbations.”
Foveation-based Mechanisms Alleviate Adversarial Examples (2015, background): “Also, note that for each CNN architecture the perturbation is different.”
Disentangling Adversarial Robustness and Generalization (2018, background): “Generalization: Research also includes independent benchmarks of attacks and defenses [13, 15, 5, 6, 85], their properties [57, 84], as well as theoretical questions [34, 42, 23, 99, 27, 88, 102, 104].”
Towards Robust Deep Neural Networks with BANG (2016, background): “As hypothesized in [4, 2], adversarial instability exists due to DNNs acting as high-dimensional linear classifiers that allow even imperceptibly small, well-aligned perturbations applied to inputs to spread among higher dimensions and radically change the outputs.”
Fine-grained Recognition in the Noisy Wild: Sensitivity Analysis of Convolutional Neural Networks Approaches (2016, background): “Furthermore, the study of [3] focuses on adversarial noise rather than random noise.”
6G Networks: Beyond Shannon Towards Semantic and Goal-Oriented Communications (2020, background): “current deep learning systems may sometimes be so brittle and easy to fool despite their uncanny power [68]: They search for correlations in data, rather than meaning, but meaning is much more than correlation.”
Adversarial Examples on Object Recognition (2020, background): “tly approximate the input distribution. However, more evidence is needed to support this conjecture for models with high capacity. Some publications suggest there are limits to adversarial robustness [19, 49, 59] and even that sensitivity to adversarial examples can not be removed [59]. Such a conse”
When Machine Learning Meets Privacy (2020, background): “ep learning itself. Some researchers recently found that there are limitations to deep learning. Specifically, “it is proved to be vulnerable to some well-designed inputs termed adversarial examples” [36, 138]. Szegedy et al. [152] first discovered that the superposition ACM Comput. Surv., Vol. 1, N”
A Theoretical Framework for Robustness of (Deep) Classifiers against Adversarial Examples (2016, background): “Subsequent papers (Fawzi et al., 2015; Papernot et al., 2015a; Nguyen et al., 2015) have explored other ways to explore adversarial examples for DNN (details in Section 2.1).”