Transferable Perturbations of Deep Feature Distributions

Nathan Inkawhich, Kevin J. Liang, Lawrence Carin, Yiran Chen

Abstract

Almost all current adversarial attacks of CNN classifiers rely on information derived from the output layer of the network. This work presents a new adversarial attack based on the modeling and exploitation of class-wise and layer-wise deep feature distributions. We achieve state-of-the-art targeted blackbox transfer-based attack results for undefended ImageNet models. Further, we place a priority on explainability and interpretability of the attacking process. Our methodology affords an analysis of how adversarial attacks change the intermediate feature distributions of CNNs, as well as a measure of layer-wise and class-wise feature distributional separability/entanglement. We also conceptualize a transition from task/data-specific to model-specific features within a CNN architecture that directly impacts the transferability of adversarial examples.

Introduction

Adversarial attacks—adding imperceptible perturbations to images to mislead deep learning models—face a significant challenge in "transferability." An attack crafted for a specific "whitebox" model often fails against a different "blackbox" target. The authors argue this is because standard attacks focus on the final logit layer, which is highly model-specific. To create more transferable attacks, they propose targeting intermediate features that represent the general "essence" of a class before the network applies architecture-specific processing.

Methodology: Feature Distribution Attack (FDA)

The core innovation is the Feature Distribution Attack (FDA). Instead of using the final classification output, FDA targets intermediate layers using a two-stage process:

  1. Auxiliary Modeling: Small binary classifiers ($g_{l,c}$) are trained for specific layers ($l$) and classes ($c$) to learn the distribution of intermediate features for that class.
  2. Adversarial Perturbation: The input image is perturbed so that its intermediate features maximize the probability of being classified as the target class by the auxiliary model.
The FDA framework: (Top) Training auxiliary classifiers on intermediate features. (Bottom) Generating adversarial examples by maximizing target class probability through these classifiers.

The researchers developed several variants, with FDA+fd (Feature Disruption) proving most effective. This variant maximizes target class probability while simultaneously forcing the perturbed features to be as different as possible from the original ones, effectively "breaking" the source representation.

Identifying the "Sweet Spot" for Transfer

A key insight of this work is the discovery of an optimal transition point within CNNs for generating attacks. Early layers are too general (task-agnostic), while final layers are overly model-specific. The most transferable attacks emerge from middle-to-late layers, where the network has captured robust class concepts but hasn't yet overfitted them to its own internal logic.

Performance metrics (Error and Targeted Success Rate) across different layer depths for various model transfers. The peaks in tSuc highlight the optimal layers for transferable attacks.

As shown in the experimental results, FDA+fd significantly outperforms standard baselines like PGD (tpgd) and Momentum-based attacks (tmim). In a transfer from DenseNet-121 to ResNet-50, where standard attacks often yield near-zero success, FDA+fd achieves a targeted success rate (tSuc) of approximately 30.9%.

Conclusion

The Feature Distribution Attack demonstrates that targeting intermediate feature distributions is a powerful strategy for creating transferable adversarial examples. By shifting the focus from final outputs to internal representations, the researchers achieved state-of-the-art results and provided new tools for interpreting how CNNs process class-specific information. Future work will involve testing these attacks against defended models and automating the selection of optimal layers for disruption.

Summary updated

Abstract

Almost all current adversarial attacks of CNN classifiers rely on information derived from the output layer of the network. This work presents a new adversarial attack based on the modeling and exploitation of class-wise and layer-wise deep feature distributions. We achieve state-of-the-art targeted blackbox transfer-based attack results for undefended ImageNet models. Further, we place a priority on explainability and interpretability of the attacking process. Our methodology affords an analysis of how adversarial attacks change the intermediate feature distributions of CNNs, as well as a measure of layer-wise and class-wise feature distributional separability/entanglement. We also conceptualize a transition from task/data-specific to model-specific features within a CNN architecture that directly impacts the transferability of adversarial examples.
BibTeX
@inproceedings{inkawhich2020transferable,
  author = {Nathan Inkawhich and Kevin J. Liang and Lawrence Carin and Yiran Chen},
  title = {{Transferable Perturbations of Deep Feature Distributions}},
  booktitle = {International Conference on Learning Representations},
  year = {2020},
  url = {https://openreview.net/forum?id=POFpQaVawm},
}
Figure