Universal, transferable and targeted adversarial attacks

Junde Wu, Rao FuarXiv 2019 · OpenReview

TL;DR

This paper introduces a novel method for generating adversarial attacks that are simultaneously universal (one attack works for all images), transferable (works across different models), and targeted (forces a specific misclassification). They achieve this by first generating "low-frequency fooling images" that are inherently more transferable and then using a conditional generative network to transform source images into adversarial examples that inherit these properties.

What Problem Does This Paper Address?

Deep neural networks (DNNs) have achieved remarkable success, but they are also surprisingly vulnerable to adversarial examples. These are inputs that have been subtly modified, often imperceptibly to humans, to cause the DNN to make incorrect predictions. While various types of adversarial attacks exist (targeted vs. non-targeted, universal vs. individual, transferable vs. non-transferable), achieving all these desirable properties simultaneously has been a significant challenge. Specifically, targeted attacks are generally harder to make transferable and universal, and existing methods often require optimizing an attack for each individual image, making them computationally expensive and impractical for real-world black-box scenarios. This paper aims to bridge this gap by proposing a method that can generate attacks that are universal, transferable, and targeted.

Methods

The core of the proposed method involves two main stages:

  1. Low-Frequency Fooling Images (Ilf): The authors hypothesize that "low-frequency" fooling images (images that are meaningless to humans but fool classifiers with high confidence) are more transferable than "high-frequency" ones. They explore several techniques to generate these low-frequency images, including:
  2. Transformation Robustness (TR): Applying small transformations (rotation, scaling, jitter) to the image before optimization.
  3. Decorrelation (DR): Using gradient descent in the Fourier basis to decorrelate pixel relationships.
  4. Gradient optimized Compositional Pattern Producing Network (Gradient-CPPN): Using a CPPN to generate images, where the network's structure influences the frequency.
  5. They found that Gradient-CPPN produced the best low-frequency fooling images.
  6. Fooling Transfer Net (FTN): This is a conditional image generation network (an encoder-decoder architecture) designed to learn a universal mapping from source images to adversarial examples. FTN takes a source image and a low-frequency fooling image as input. It uses the high-level representations of the fooling image (extracted from a pre-trained classifier) to guide the generation process. The goal is to shift the source image's distribution to inherit the targeted and transferable properties of the fooling image, while maintaining the original image's content. FTN is trained using a combination of:
  7. Content Loss (Lc): Using Structural Similarity Index Measure (SSIM) to preserve pixel-level similarity.
  8. Representation Loss (Lrep): Using Maximum Mean Discrepancy (MMD) to match the high-level representations of the generated adversarial examples with those of the fooling images.
  9. Total Variance Loss (Ltv): To penalize reconstruction noise.

Datasets

  • Training: ILSVRC 2017 classification training set.
  • Evaluation: The paper tests transferability on various pre-trained classification models (Inception-v3, ResNet-18, ResNet-50, DenseNet) and a black-box image classification server (Clarifai.com).

Key Insights and Ideas

  • Low-frequency is transferable: The central hypothesis that low-frequency fooling images are more transferable than high-frequency ones is empirically supported. This suggests that attacks targeting more "natural" image manifolds are more robust across different models.
  • Decoupling perturbation from boundary curvature: The authors propose that their method decouples the adversarial perturbation from the specific curvature of a single model's decision boundary. By focusing on properties of low-frequency fooling images, which are less sensitive to model-specific boundaries, the generated attacks become more transferable.
  • Universal mapping for targeted attacks: The FTN network learns a single mapping that can transform any source image into a targeted adversarial example, achieving universality for targeted attacks, which was previously difficult.

Key Results

  • The proposed method (FTN) achieves significantly higher transfer success rates compared to other methods, including individual attacks (FG, JSMA, DeepFool, C&W) and other universal attack methods.
  • For example, FTN achieved a transfer success rate of 98% on Inception-v3 and 94% on Clarifai.com, while other methods like C&W achieved only 2-3%.
  • The generated adversarial examples exhibit low distortion (low RMSD values), meaning they are visually similar to the original images.
  • The low-frequency fooling images generated by Gradient-CPPN performed best among the explored methods for generating transferable fooling images.

Limitations and Future Work

  • Limitations: The paper acknowledges that while their method achieves high transferability, the "meaning" of low-frequency attacks and why they are more transferable is still an area for further discussion. The specific choice of representation layers for supervision in FTN can impact performance.
  • Future Work: The authors suggest exploring the use of neural network discriminators to measure high-level representation similarity and analyzing fooling image clusters on the manifold to further support their hypotheses. They also mention that targeted attacks on other classes can be found on their project homepage.

Summary updated

Abstract

Deep Neural Networks have been found vulnerable re-cently. A kind of well-designed inputs, which called adver-sarial examples, can lead the networks to make incorrectpredictions. Depending on the different scenarios, goalsand capabilities, the difficulties of the attacks are different.For example, a targeted attack is more difficult than a non-targeted attack, a universal attack is more difficult than anon-universal attack, a transferable attack is more difficultthan a nontransferable one. The question is: Is there existan attack that can meet all these requirements? In this pa-per, we answer this question by producing a kind of attacksunder these conditions. We learn a universal mapping tomap the sources to the adversarial examples. These exam-ples can fool classification networks to classify all of theminto one targeted class, and also have strong transferability.Our code is released at: xxxxx.
BibTeX
@misc{wu2019universal,
  author = {Junde Wu and Rao Fu},
  title = {{Universal, transferable and targeted adversarial attacks}},
  year = {2019},
  eprint = {1908.11332},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1908.11332},
}
Figure