Abstract: The susceptibility of deep learning models to adversarial perturbations has stirred renewed attention in adversarial examples resulting in a number of attacks. However, most of these attacks fail to encompass a large spectrum of adversarial perturbations that are imperceptible to humans. In this paper, we present localized uncertainty attacks, a novel class of threat models against deterministic and stochastic classifiers. Under this threat model, we create adversarial examples by perturbing only regions in the inputs where a classifier is uncertain. To find such regions, we utilize the predictive uncertainty of the classifier when the classifier is stochastic or, we learn a surrogate model to amortize the uncertainty when it is deterministic. Unlike $\ell_p$ ball or functional attacks which perturb inputs indiscriminately, our targeted changes can be less perceptible. When considered under our threat model, these attacks still produce strong adversarial examples; with the examples retaining a greater degree of similarity with the inputs.
Introduction
Deep learning models are notoriously fragile. By adding tiny, often invisible perturbations to an input—known as adversarial examples—attackers can trick a model into misclassifying a "Panda" as a "Gibbon." However, traditional attacks (like those using $\ell_p$ balls) usually perturb every single pixel in an image indiscriminately. While mathematically "small," these global changes often create a hazy, noisy, or "unnatural" look that a human can easily spot.
This research proposes that we do not need to change the whole image; we just need to find the regions where the model is "poorly constrained"—essentially, the areas where the model is guessing or extrapolating into the unknown. By identifying and perturbing only the specific regions where a classifier is already uncertain, these attacks achieve high success rates while remaining significantly more imperceptible to the human eye.
Methodology
The core innovation is a two-pronged approach to localizing model vulnerability through a specialized pipeline.
The Mask Model
The authors train a secondary neural network, the "mask model," which takes an image as input and outputs a binary mask. This mask identifies specific pixels or regions to be perturbed. The goal of this mask is to maximize the epistemic uncertainty (the uncertainty resulting from a lack of data/knowledge) of the target classifier. The final attack combines this by optimizing a loss function that balances three things: making the model fail, targeting uncertain regions, and keeping the mask "sparse" (using as few pixels as possible).
Amortized Uncertainty
Measuring uncertainty is straightforward for stochastic (Bayesian) models but difficult for standard deterministic ones. To solve this, the authors train a surrogate uncertainty model. This surrogate learns to estimate the predictive uncertainty of a deterministic classifier, allowing the attack to work against almost any standard architecture by amortizing the uncertainty calculation.
Experimental Evaluation
The researchers evaluated their approach on standard computer vision benchmarks including MNIST, CIFAR-10, STL-10, and ImageNet. The localized approach results in much lower Learned Perceptual Image Patch Similarity (LPIPS) scores, meaning the images look much more "natural" to humans despite being adversarial.
Key Results
- High Success, Low Footprint: On CIFAR-10, the "UNC.
- delta" attack achieved a 0% accuracy (total success) against undefended models while perturbing significantly fewer pixels than vanilla baselines.
- Improved Imperceptibility: In STL-10 experiments, the localized attack reduced perceptual distortion (LPIPS) by nearly 50% compared to standard additive attacks.
- Transferability: Adversarial examples generated against deterministic models using the surrogate uncertainty method successfully transferred to stochastic (Bayesian) models, proving the underlying vulnerabilities are consistent across architectures.
Discussion and Future Work
By targeting regions of high uncertainty, the attack essentially "pushes" the model off a cliff it was already standing near. This supports the hypothesis that adversarial examples often lie "off the manifold" of natural data; regions where a model makes unconstrained extrapolations are holes in its logic that are easy to exploit.
While highly effective, the current method is a "white-box" attack, requiring access to the model's internal gradients. Future work may explore extending this to "black-box" scenarios and applying the framework to other data types like text or audio, where localized vulnerabilities may also exist.
Summary updated