Are Generative Classifiers More Robust to Adversarial Attacks?

By Yingzhen Li, John Bradshaw, Yash Sharma. 2018

87 citations recorded by the source corpus.

There is a rising interest in studying the robustness of deep neural network classifiers against adversaries, with both advanced attack and defence techniques being actively developed. However, most recent work focuses on discriminative classifiers which only models the conditional distribution of the labels given the inputs. In this abstract we propose deep Bayes classifier that improves the classical naive Bayes with conditional deep generative models, and verifies its robustness against a number of existing attacks. We further developed a detection method for adversarial examples based on conditional deep generative models. Our initial results on MNIST suggest that deep Bayes classifiers might be more robust when compared with deep discriminative classifiers, and the proposed detection method achieves high detection rates against two commonly used attacks.

Subfields: Artificial Intelligence, Generative Models, Safety, Robustness & Interpretability, Signal Processing

How researchers see this paper

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

In Artificial Intelligence

In Generative Models

In Safety, Robustness & Interpretability

In Signal Processing

Citations in the corpus

A Wholistic View of Continual Learning with Deep Neural Networks: Forgotten Lessons and the Bridge to Active and Open World Learning (2020, background): “Reinforced Continual Learning (RCL) (Xu and Zhu, 2018) or Learn-to-Grow (Li et al., 2019a) further attempt to overcome the challenge of ﬁnding suitable loss cut-oﬀs and cast dynamic unit addition into a meta-learning framework in order to separate the learning of the network structure and…”
Anomalous Instance Detection in Deep Learning: A Survey (2020, methodology): “In [72], the method improves the naive Bayes used in many generative classifiers by combining it with variational auto-encoder.”
Conditional Generative Models are not Robust (2019, methodology): “In [Li et al., 2018] the authors report difficulties training a high accuracy conditional generative model on CIFAR10, and resort to evaluation on a 2-class classification problem derived from CIFAR10.”
Advances in adversarial attacks and defenses in computer vision: A survey (2021, unknown): “These defenses include [158], [159], [160], [161], [162], [163], [164], [165], [166], [167], [168], [169] and [170].”
Understanding the Limitations of Conditional Generative Models (2019, background): “Recent work (Song et al., 2018; Frosst et al., 2018; Li et al., 2018) showed that a generative model can detect and defend adversarial attacks.”
On Adaptive Attacks to Adversarial Example Defenses (2020, methodology): “As we saw previously with the generative model defense (Li et al., 2019) in Section 7, attacks on complex defenses do not have to be complex once the weak links are identified.”