Robust sensible adversarial learning of deep neural networks for image classification

Jungeum Kim, Xiao Wang

TL;DR

This paper introduces "sensible adversarial learning," a new approach to train deep neural networks that are robust to adversarial attacks without sacrificing natural accuracy. They propose a novel algorithm, SENSE-AT, which effectively balances these competing goals, even for models with limited capacity.

What Problem Does This Paper Address?

Deep neural networks (DNNs) have achieved remarkable success in tasks like image classification. However, they are notoriously vulnerable to adversarial attacks: tiny, imperceptible changes to input data can cause DNNs to misclassify them with high confidence. This "adversarial vulnerability" poses a significant threat to the trustworthiness and safety of AI systems. A common challenge in defending against these attacks is the "robustness-accuracy trade-off": improving a model's robustness often leads to a decrease in its accuracy on clean, unperturbed data. This paper aims to resolve this trade-off by introducing a new framework for adversarial learning.

Methods

The core of this work is the concept of a "sensible adversary." Unlike traditional adversarial attacks that aim to maximize misclassification within a fixed perturbation budget (e.g., an epsilon-ball), a sensible adversary is defined to avoid crossing the Bayes decision boundary. This means it only generates adversarial examples that would still be correctly classified by the ideal Bayes classifier.

The authors theoretically prove that the Bayes classifier is the most robust under this sensible adversarial framework. They then propose a novel training algorithm called SENSE-AT (Sensible Adversarial Training). SENSE-AT uses a technique called "implicit loss truncation" to handle the challenges of optimizing for both natural accuracy and robustness. This involves a modified loss function that effectively prioritizes natural accuracy when the model has sufficient capacity, and robust accuracy when necessary, without requiring explicit truncation of the loss function itself. The algorithm is designed to be efficient and does not require additional networks or datasets.

Datasets

  • MNIST: A standard dataset of handwritten digits (0-9) used for image classification.
  • CIFAR10: A dataset of 60,000 32x32 color images in 10 classes, commonly used for image classification benchmarks.
  • Three Clusters Dataset: A synthetic dataset introduced by the authors to specifically analyze the trade-off between robustness and accuracy, particularly in scenarios with limited model capacity.

Key Insights and Ideas

  • The Bayes Classifier is "Sensibly" Robust: The paper theoretically demonstrates that the ideal Bayes classifier is inherently robust to "sensible" adversarial attacks, meaning attacks that don't cross the true decision boundary. This suggests that by staying close to the Bayes decision boundary, robustness can be achieved without sacrificing natural accuracy.
  • Sensible Adversary vs. Standard Adversary: The key difference lies in the goal. Standard adversaries aim to fool the model within a fixed perturbation budget. Sensible adversaries aim to fool the model without crossing the true class boundary, effectively allowing for more "gullibility" near the class border if it preserves natural accuracy.
  • Implicit Loss Truncation for Practicality: Directly optimizing truncated loss functions can be challenging due to zero gradients. SENSE-AT's "implicit loss truncation" cleverly reformulates the problem to avoid this, making optimization more tractable and efficient.
  • Robustness Without Collapse for Limited Capacity: A significant finding is that SENSE-AT can train robust models even when the model capacity is insufficient for standard adversarial training (R-AT) to avoid collapsing into a constant function. This is achieved by prioritizing natural accuracy when robustness would otherwise lead to a collapse.

Key Results

  • High Natural and Robust Accuracy: On CIFAR10, SENSE-AT achieved over 90% natural test accuracy while maintaining comparable adversarial test accuracy against various attacks.
  • No Collapse with Limited Capacity: Experiments showed that SENSE-AT models did not collapse even with small model capacities, unlike R-AT which often collapsed.
  • Insensitivity to Hyperparameter 'c': The main hyperparameter 'c' (controlling the trade-off) was found to be not overly sensitive, promoting the practicality of the method.
  • Competitive Performance: SENSE-AT demonstrated competitive results compared to state-of-the-art adversarial training methods like TRADES, MART, and IAAT.

Limitations and Future Work

  • Generalization Error Bounds: The authors note that a remaining theoretical problem is to develop generalization error bounds for sensible adversarial training to formally justify the empirical performance.
  • Sample Size: The paper did not explicitly address the issue of limited sample size, which is an area for future theoretical investigation.
  • Transfer Attacks: While SENSE-AT showed comparable performance against transfer attacks, further research could explore its specific strengths and weaknesses in this domain.

Summary updated

Abstract

The idea of robustness is central and critical to modern statistical analysis. However, despite the recent advances of deep neural networks (DNNs), many studies have shown that DNNs are vulnerable to adversarial attacks. Making imperceptible changes to an image can cause DNN models to make the wrong classification with high confidence, such as classifying a benign mole as a malignant tumor and a stop sign as a speed limit sign. The trade-off between robustness and standard accuracy is common for DNN models. In this paper we introduce sensible adversarial learning and demonstrate the synergistic effect between pursuits of standard natural accuracy and robustness. Specifically, we define a sensible adversary, which is useful for learning a robust model, while keeping high natural accuracy. We theoretically establish that the Bayes classifier is the most robust multiclass classifier with the 0−1 loss under sensible adversarial learning. We propose a novel and efficient algorithm that trains a robust model using implicit loss truncation. We apply sensible adversarial learning for large-scale image classification to a handwritten digital image dataset, called MNIST, and an object recognition colored image dataset, called CIFAR10. We have performed an extensive comparative study to compare our method with other competitive methods. Our experiments empirically demonstrate that our method is not sensitive to its hyperparameter and does not collapse even with a small model capacity while promoting robustness against various attacks and keeping high natural accuracy. The sensible adversarial learning software is available as a Python package at https://github.com/JungeumKim/SENSE.
BibTeX
@misc{kim2022robust,
  author = {Jungeum Kim and Xiao Wang},
  title = {{Robust sensible adversarial learning of deep neural networks for image classification}},
  year = {2022},
  eprint = {2205.10457},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2205.10457},
}
Figure