Making Adversarial Examples More Transferable and Indistinguishable

Junhua Zou, Yexin Duan, Boyu Li, Wu Zhang, Yu Pan, Zhisong PanAAAI 2022 · DOI

Abstract

Adversarial examples pose a significant threat to deep neural networks (DNNs), yet creating attacks that are both highly transferable to unknown models and visually indistinguishable to humans remains a challenge. This paper introduces AI-FGTM (Adam Iterative Fast Gradient Tanh Method), a novel approach that leverages the Adam optimizer and a smooth tanh activation function to replace the traditional sign-based gradient updates. Our method preserves fine-grained gradient information, achieving state-of-the-art transferability while reducing perturbation size by nearly 20%.

Introduction

Deep neural networks are notoriously vulnerable to adversarial examples—inputs subtly modified to cause misclassification. Two primary obstacles hinder the effectiveness of these attacks in real-world scenarios:

  1. Transferability: Attacks optimized for a specific "white-box" model often fail against different "black-box" architectures.
  2. Indistinguishability: Standard methods like the Fast Gradient Sign Method (FGSM) use the sign function, which forces pixel perturbations to extreme values, creating visible, blocky artifacts.

We argue that the sign function acts as a blunt instrument, discarding critical gradient texture. By introducing AI-FGTM, we demonstrate that smoother optimization leads to more potent and stealthier adversarial perturbations.

Methodology: The AI-FGTM Framework

The proposed AI-FGTM method refines the iterative attack pipeline through four key innovations:

1. The Tanh Activation Swap

Traditional iterative attacks utilize sign(gradient) to update pixels. This binarization process is essentially a low-bit quantization that destroys high-frequency information. We replace this with the tanh function. Near zero, tanh behaves linearly ($y \approx x$), preserving small gradient variations, while its saturation at higher values prevents excessive pixel changes. This results in "noise" that mimics natural image textures.

2. Adam Optimizer Integration

While momentum-based methods are common, we adapt the Adam optimizer for the adversarial generation phase. By tracking both the first and second moments of gradients, Adam allows the attack to converge more efficiently. Remarkably, this optimization remains effective even within the extremely short horizon of 10 iterations typical for adversarial attacks.

3. Dynamic Step Size

Instead of a static learning rate, we employ a dynamic step size that increases over the course of the iterations. This allows the attack to start with subtle adjustments and "push" more aggressively as it approaches the decision boundary of the target model.

Dynamic step size curve used in AI-FGTM to accelerate convergence.

4. Refined Gaussian Blurring

In Translation-Invariant (TI) attacks, gradients are often blurred to improve robustness. We found that using smaller Gaussian kernels prevents the loss of critical high-frequency information, further enhancing the attack's success rate.

The integrated version of our approach, NI-TI-DI-AITM, combines these features with Nesterov momentum and Diverse Inputs (DI).

Experimental Evaluation

We evaluated our method using an ImageNet-compatible dataset of 1,000 images. The attack's performance was tested against a variety of architectures, including:

  • White-box models: Inception v3, Inception v4, Inception ResNet v2, and ResNet v2-101.
  • Defensive models: Classic defenses (e.g., HGD, R&P) and advanced defenses (e.g., Feature Distillation, Randomized Smoothing).

Results: Transferability and Stealth

Our comprehensive attack, NI-TI-DI-AITM, achieved an average success rate of 89.3% against classic defenses and 82.7% against advanced defenses, significantly outperforming existing state-of-the-art methods.

Crucially, this increase in power does not come at the cost of visibility. AI-FGTM reduced the mean perturbation size by approximately 20% compared to baseline iterative attacks. As shown in the visual comparisons below, the perturbations generated by our method are significantly less noticeable than those from standard DIM or TI-DIM methods.

Visual comparison of clean images and adversarial examples generated by various methods. AI-FGTM (AITM) variants show fewer visible artifacts.

Conclusion

AI-FGTM represents a shift from "blunt" gradient manipulation to refined optimization in adversarial machine learning. By preserving gradient information through the tanh function and utilizing the Adam optimizer, we achieve a superior trade-off between attack success and image quality. This work establishes a new, stronger baseline for evaluating the robustness of future defensive AI systems.

Summary updated

BibTeX
@inproceedings{zou2022making,
  author = {Junhua Zou and Yexin Duan and Boyu Li and Wu Zhang and Yu Pan and Zhisong Pan},
  title = {{Making Adversarial Examples More Transferable and Indistinguishable}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i3.20279},
}
Figure