AI-GAN: Attack-Inspired Generation of Adversarial Examples

Tao Bai, Jun Zhao, Jinlin Zhu, Shoudong Han, Jiefeng Chen, Bo Li

Abstract

AI-GAN is a novel generative framework designed to produce high-quality adversarial examples across multiple target classes in a single forward pass. By jointly training a generator, a discriminator, and an internal "attacker," the framework achieves state-of-the-art attack success rates. AI-GAN is orders of magnitude faster than traditional optimization-based methods and demonstrates superior scalability, successfully attacking complex datasets like CIFAR-100.

Introduction

Deep Neural Networks (DNNs) have shown remarkable performance in various tasks but remain notoriously fragile when faced with adversarial examples—input data modified by tiny, human-imperceptible perturbations that cause the model to make confident, incorrect predictions.

Current methods for generating these attacks generally fall into two categories: optimization-based and generative-based. Optimization-based attacks, such as C&W and PGD, are highly effective but computationally expensive, often requiring thousands of iterations for a single image. Generative-based attacks, like AdvGAN, offer faster inference but typically struggle to scale beyond simple datasets and require retraining the entire model to change the target class of the attack. AI-GAN addresses these limitations by providing a fast, flexible, and scalable framework for multi-target adversarial generation.

The AI-GAN Framework

The core innovation of the AI-GAN (Attack-Inspired GAN) architecture is the end-to-end joint training of four interacting components:

The AI-GAN architecture featuring a generator, discriminator, and an internal attacker.
  1. The Generator ($G$): Unlike standard GANs that start from noise, this generator takes a clean image and a specific "target class" label as input. It utilizes a pre-trained encoder to extract robust features, which is essential for scaling to complex images.
  2. The Discriminator ($D$): This is a "two-headed" critic. One head performs the standard GAN task of distinguishing real images from perturbed ones, while the other acts as an auxiliary classifier to ensure the perturbations actually look like the target class to a model.
  3. The Attacker ($A$): This is the "Attack-Inspired" secret sauce. The authors include an internal attacker that generates its own adversarial examples to train the discriminator. This forces the discriminator to become a much tougher critic, which in turn pushes the generator to produce more sophisticated and effective perturbations.
  4. The Target Classifier ($C$): The "victim" model that the generator is attempting to fool.

By training these jointly, the generator learns a general mapping of how to shift any image into any target class within a defined perturbation bound ($L_\infty$).

Experimental Evaluation

The researchers evaluated AI-GAN on three standard computer vision benchmarks: MNIST, CIFAR-10, and CIFAR-100. A key highlight is the framework's ability to scale to CIFAR-100, achieving an 87.76% success rate—a significant milestone for generative adversarial models which have previously struggled with such complexity.

Adversarial examples generated by AI-GAN on CIFAR-10, demonstrating multi-target capability.

In terms of efficiency, AI-GAN is incredibly fast, generating adversarial examples in less than 0.01 seconds. This is a massive improvement over optimization-based methods like the C&W attack, which can take over 3 hours for similar tasks. Furthermore, AI-GAN achieves near-perfect success rates on simpler datasets (99.14% on MNIST) and maintains high performance even against models protected by adversarial training, proving that its learned perturbations are more "meaningful" and harder to defend against.

Conclusion and Future Work

AI-GAN represents a significant advancement in adversarial research by providing a "many-to-many" generation capability where a single trained generator can target multiple classes. While the current study focuses on the $L_\infty$ norm and 32x32 resolution images, the success on CIFAR-100 paves the way for scaling these techniques to high-resolution ImageNet-scale data and exploring different perturbation bounds in future research.

Summary updated

Abstract

Deep neural networks (DNNs) are vulnerable to adversarial examples, which are crafted by adding imperceptible perturbations to inputs. Recently different attacks and strategies have been proposed, but how to generate adversarial examples perceptually realistic and more efficiently remains unsolved. This paper proposes a novel framework called Attack-Inspired GAN (AI-GAN), where a generator, a discriminator, and an attacker are trained jointly. Once trained, it can generate adversarial perturbations efficiently given input images and target classes. Through extensive experiments on several popular datasets e.g., MNIST and CFAR-10, AI-GAN achieves high attack success rates and reduces generation time significantly in various settings. Moreover, for the first time, AI-GAN successfully scales to complicated datasets e.g., CFAR-100 with around 90% success rates among all classes.
BibTeX
@misc{bai2020ai,
  author = {Tao Bai and Jun Zhao and Jinlin Zhu and Shoudong Han and Jiefeng Chen and Bo Li},
  title = {{AI-GAN: Attack-Inspired Generation of Adversarial Examples}},
  year = {2020},
  eprint = {2002.02196},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2002.02196},
}
Figure