Searching for Robust Neural Architectures via Comprehensive and Reliable Evaluation

Jialiang Sun, Tingsong Jiang, Chao Li, Weien Zhou, Xiaoya Zhang, Wen Yao, Xiaoqian ChenarXiv 2022 · OpenReview

TL;DR

This paper introduces A³D, a novel platform that automates the search for both robust neural network architectures and efficient adversarial attacks. By unifying these two crucial aspects of DNN security, A³D aims to overcome the limitations of existing tools that can only evaluate, not optimize, these components, leading to more secure and reliable deep learning models.

What Problem Does This Paper Address?

Deep neural networks (DNNs) are increasingly deployed in safety-critical applications like self-driving cars and facial recognition systems. However, their vulnerability to adversarial examples (AEs) – subtly perturbed inputs that cause misclassifications – poses a significant security risk. While many tools exist to evaluate DNN robustness against adversarial attacks, they often fall short because they cannot simultaneously optimize DNN architectures and the adversarial attacks themselves. This disconnect means that a robust architecture might not be tested against its strongest possible attack, or an efficient attack might not be developed with the best possible target architecture in mind. This paper addresses this gap by proposing a unified framework, A³D, that automates the search for both robust architectures and potent adversarial attacks, enabling a more comprehensive and effective approach to DNN security.

Methods

The A³D framework is built upon two core AutoML components:

  • Auto Adversarial Defense (Neural Architecture Search - NAS): This module employs various NAS techniques (differentiable and non-differentiable) to find robust neural network architectures. Crucially, it incorporates a comprehensive suite of four types of robustness evaluations:
  • Adversarial Noises: Standard adversarial attacks like FGSM and PGD.
  • Natural Noises: Real-world corruptions like brightness, contrast, fog, and frost (using datasets like CIFAR10-C and ImageNet-C).
  • System Noises: Imperfections arising from system implementations (e.g., image decoding and resizing).
  • Quantified Metrics: Mathematical measures like the Frobenius norms of Jacobian and Hessian matrices, which assess the landscape smoothness of DNN predictions. The framework allows these evaluations to be combined into loss functions for NAS, guiding the search towards architectures that are resilient to diverse threats.

  • Auto Adversarial Attack (Evolutionary Search): This module uses evolutionary algorithms (single-objective and multi-objective) to search for efficient adversarial attacks. It defines a mathematical model for adversarial attacks that considers attacker operations, loss functions, magnitudes, and iteration numbers. This allows for the discovery of attacks that achieve a desired level of robustness degradation with minimal computational cost.

The key innovation is the unified framework that allows these two components to interact. The searched efficient attacks can serve as new, more challenging robustness evaluations for NAS, while the searched robust architectures can act as threat models for developing stronger adversarial attacks.

Datasets

The paper utilizes standard image classification datasets for its experiments:

  • CIFAR10
  • CIFAR100
  • ImageNet

Additionally, it leverages datasets for natural noise evaluation:

  • CIFAR10-C
  • ImageNet-C

Key Insights and Ideas

  • Unified Search is Crucial: The paper strongly suggests that optimizing architectures and attacks in isolation is suboptimal. A joint search process, where each component informs the other, leads to more effective security solutions.
  • Diverse Robustness Evaluations Matter: Relying solely on traditional adversarial attacks (like FGSM/PGD) for evaluating robustness is insufficient. Incorporating natural and system noises, along with quantified metrics, provides a more holistic and realistic assessment of a model's resilience.
  • Quantified Metrics are Informative but Not Definitive: While metrics like Jacobian and Hessian can help identify smoother landscapes (indicating robustness), the paper shows that the lowest quantified metric doesn't always translate to the highest robustness against actual attacks.
  • Trade-offs Exist: The search for efficient attacks highlights the inherent trade-off between attack strength (robust accuracy) and computational cost (time). Multi-objective optimization is key to finding a desirable balance.
  • NAS Can Be Guided by Stronger Attacks: Using the output of an automated attack search as an evaluation metric for NAS can lead to more robust architectures than using simpler, manually-designed attacks.

Key Results

  • The A³D framework successfully demonstrates its ability to find both more robust neural network architectures and more efficient adversarial attacks across CIFAR10, CIFAR100, and ImageNet datasets.
  • Experiments show that the searched adversarial attacks often outperform manually-designed attacks in terms of achieving lower robust accuracy. For instance, in some cases, the robust accuracy dropped significantly more with searched attacks compared to standard ones.
  • The paper showcases that different NAS methods (e.g., DARTS, PC-DARTS, NASP, FairDARTS, SmoothDARTS, Random Search, DE) yield architectures with varying robustness characteristics depending on the evaluation metric used during the search.
  • The unified framework's ability to leverage searched attacks as evaluations for NAS and vice-versa is demonstrated, leading to improved overall security. For example, using a searched attack as an evaluation metric for NAS resulted in a higher robust accuracy (49.4%) compared to using a standard FGSM attack (41.6%).

Limitations and Future Work

  • Limited Evolutionary Algorithms/NAS Methods: The authors acknowledge that A³D includes "some" evolutionary algorithms and NAS methods, implying that not all state-of-the-art techniques might be implemented.
  • Focus on Search Strategies and Evaluations: The current work prioritizes the performance of different search strategies and evaluation metrics. Future work could explore the impact of different search spaces.
  • Broader Impact: The paper suggests that A³D can serve as a benchmark and toolkit for researchers. Future work could expand its applicability to other domains beyond image classification.
  • More Robust Architectures: The paper aims to find more robust architectures under searched stronger attacks, and conversely, find more efficient attacks under searched architectures, indicating a continuous improvement loop.

Summary updated

Abstract

The robustness of deep neural networks (DNN) models has attracted increasing attention due to the urgent need for security in many applications. Numerous existing open-sourced tools or platforms are developed to evaluate the robustness of DNN models by ensembling the majority of adversarial attack or defense algorithms. Unfortunately, current platforms do not possess the ability to optimize the architectures of DNN models or the configuration of adversarial attacks to further enhance the robustness of models or the performance of adversarial attacks. To alleviate these problems, in this paper, we first propose a novel platform called auto adversarial attack and defense ($A^{3}D$), which can help search for robust neural network architectures and efficient adversarial attacks. In $A^{3}D$, we employ multiple neural architecture search methods, which consider different robustness evaluation metrics, including four types of noises: adversarial noise, natural noise, system noise, and quantified metrics, resulting in finding robust architectures. Besides, we propose a mathematical model for auto adversarial attack, and provide multiple optimization algorithms to search for efficient adversarial attacks. In addition, we combine auto adversarial attack and defense together to form a unified framework. Among auto adversarial defense, the searched efficient attack can be used as the new robustness evaluation to further enhance the robustness. In auto adversarial attack, the searched robust architectures can be utilized as the threat model to help find stronger adversarial attacks. Experiments on CIFAR10, CIFAR100, and ImageNet datasets demonstrate the feasibility and effectiveness of the proposed platform, which can also provide a benchmark and toolkit for researchers in the application of automated machine learning in evaluating and improving the DNN model robustnesses.
BibTeX
@misc{sun2022searching,
  author = {Jialiang Sun and Tingsong Jiang and Chao Li and Weien Zhou and Xiaoya Zhang and Wen Yao and Xiaoqian Chen},
  title = {{Searching for Robust Neural Architectures via Comprehensive and Reliable Evaluation}},
  year = {2022},
  eprint = {2203.03128},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2203.03128},
}
Figure