Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Resilience and Security of Deep Neural Networks Against Intentional and Unintentional Perturbations: Survey and Research Challenges

Sazzad Sayyed, Milin Zhang, Shahriar Rifat, Ananthram Swami, Michael De Lucia, Francesco Restuccia

TL;DR

This survey addresses the resilience of deep neural networks to both intentional adversarial perturbations and unintentional distributional shifts. It consolidates a unified view by linking detection and robustness approaches across input, latent, and logit spaces, covering adversarial, OOD, and open-set challenges. The authors categorize detection methods, discuss cross-domain connections, and outline critical open questions and research directions, aiming to spur collaboration between previously separate communities. The work highlights the need for scalable, explainable, and dataset-agnostic solutions, and calls for benchmarks that reflect real-world conditions and constraints. Overall, the paper provides a comprehensive roadmap for advancing resilient and secure DNN inference in practical settings.

Abstract

In order to deploy deep neural networks (DNNs) in high-stakes scenarios, it is imperative that DNNs provide inference robust to external perturbations - both intentional and unintentional. Although the resilience of DNNs to intentional and unintentional perturbations has been widely investigated, a unified vision of these inherently intertwined problem domains is still missing. In this work, we fill this gap by providing a survey of the state of the art and highlighting the similarities of the proposed approaches.We also analyze the research challenges that need to be addressed to deploy resilient and secure DNNs. As there has not been any such survey connecting the resilience of DNNs to intentional and unintentional perturbations, we believe this work can help advance the frontier in both domains by enabling the exchange of ideas between the two communities.

Resilience and Security of Deep Neural Networks Against Intentional and Unintentional Perturbations: Survey and Research Challenges

TL;DR

This survey addresses the resilience of deep neural networks to both intentional adversarial perturbations and unintentional distributional shifts. It consolidates a unified view by linking detection and robustness approaches across input, latent, and logit spaces, covering adversarial, OOD, and open-set challenges. The authors categorize detection methods, discuss cross-domain connections, and outline critical open questions and research directions, aiming to spur collaboration between previously separate communities. The work highlights the need for scalable, explainable, and dataset-agnostic solutions, and calls for benchmarks that reflect real-world conditions and constraints. Overall, the paper provides a comprehensive roadmap for advancing resilient and secure DNN inference in practical settings.

Abstract

In order to deploy deep neural networks (DNNs) in high-stakes scenarios, it is imperative that DNNs provide inference robust to external perturbations - both intentional and unintentional. Although the resilience of DNNs to intentional and unintentional perturbations has been widely investigated, a unified vision of these inherently intertwined problem domains is still missing. In this work, we fill this gap by providing a survey of the state of the art and highlighting the similarities of the proposed approaches.We also analyze the research challenges that need to be addressed to deploy resilient and secure DNNs. As there has not been any such survey connecting the resilience of DNNs to intentional and unintentional perturbations, we believe this work can help advance the frontier in both domains by enabling the exchange of ideas between the two communities.
Paper Structure (25 sections, 17 equations, 7 figures, 1 table)

This paper contains 25 sections, 17 equations, 7 figures, 1 table.

Table of Contents

  1. Introduction
  2. Motivation and Novel Contributions
  3. Vulnerabilities of Deep Neural Networks: Background and Taxonomy
  4. Intentional Perturbation: Adversarial Attack
  5. Unintentional Perturbation: Determining Out-of-Distribution Samples
  6. A Unified Vision for Intentional and Unintentional Perturbation
  7. Detection of Intentional Perturbation
  8. Use of Intermediate Representations
  9. Manifold-based detectors
  10. Interpretability-based Approaches
  11. Statistical Approaches
  12. Influence-based approaches
  13. Other Notable Approaches
  14. Detection of OOD Samples
  15. Input Space Approaches
  16. ...and 10 more sections

Figures (7)

  • Figure 1: Taxonomy of intentional perturbations.
  • Figure 2: Taxonomy of unintentional perturbation.
  • Figure 3: Overview of existing approaches for resilience.
  • Figure 4: (a): The adversarial sample $x^*$ is generated by shifting away from the 'negative' submanifold and crossing the decision boundary (black dashed line), but it remains distant from the 'positive' submanifold. (b): the 'positive' submanifold has a 'pocket' and the adversarial sample $x^*$ lies in the pocket. (c): The adversarial sample $x^*$ is close to both the decision boundary and both submanifolds feinman2017detecting.
  • Figure 5: The embedding space of a cohen2020detecting.
  • ...and 2 more figures