Abstract
Current research in deep learning robustness primarily focuses on local adversarial attacks, which evaluate model behavior in the immediate vicinity of existing data points. This paper introduces Global Adversarial Attacks, a framework designed to assess model stability across the entire input space, including regions far from the training distribution. By identifying pairs of nearly identical inputs that yield divergent model predictions, the authors reveal that standard defenses against local attacks do not guarantee global robustness. Experimental results on MNIST and CIFAR-10 show that even "hardened" models remain highly vulnerable to these global instabilities.
Introduction: Beyond Local Robustness
Most adversarial research is centered on local robustness: given a specific input $x$ with a known label $y$, can we find a nearby $x'$ that the model misclassifies? While useful, this approach is insufficient for safety-critical applications like autonomous driving. Local attacks only examine regions near the training or testing sets, leaving the vast majority of the input space as a "black box." Furthermore, they require labeled data to function.
If a model encounters a visual pattern far removed from its training data, its behavior is unpredictable. We need to determine if the model's decision boundaries are unstable throughout the input space, not just near known samples.
Methodology
The authors define a Global Adversarial Attack as the search for a pair of inputs $(x_1, x_2)$ that are visually indistinguishable (within a radius $\epsilon$) but result in different model predictions. Two primary methods are proposed:
- Alternating Gradient Attacks (G-PGD): An iterative approach where one point is fixed while the other is moved to maximize output divergence, followed by a swap. This allows the pair to "walk" through the input space toward regions of high instability.
- Extreme-Value-Guided MCMC (GEVMCMC): A stochastic approach leveraging Extreme Value Theory (EVT) and Markov Chain Monte Carlo (MCMC) sampling. This method explores the space more broadly to find diverse vulnerabilities that gradient-based methods might overlook.
To test these methods, the researchers used MNIST and CIFAR-10, augmented with an 11th class for "meaningless" random noise. This setup allows attacks to start from pure noise and observe if the model begins to perceive "meaningful" objects or sharp transitions between classes in empty space.
Results and Key Insights
The most striking finding is that local defense does not equal global robustness. Models trained using Projected Gradient Descent (PGD)—the current gold standard for local defense—remained nearly 100% vulnerable to global attacks.
- Attack Success Rate: On both benchmarks, G-PGD and GEVMCMC achieved an attack rate of nearly 100%.
- Failure of Adversarial Training: While local adversarial training reduced local attack success from 30% to 4% on MNIST, the Global PGD attack maintained a >99% success rate regardless of the defense applied.
- Vulnerability is Ubiquitous: Attacks successfully found adversarial pairs starting from both legitimate images and random noise, suggesting that neural networks possess highly fractured decision boundaries even where no real data exists.
Conclusion
This work exposes a significant gap in current robustness evaluations. While local adversarial training provides a degree of protection near the data manifold, it fails to stabilize the model's behavior globally. Future work must focus on developing Global Adversarial Training algorithms and scaling these methods to higher-resolution datasets to ensure the reliability of deep learning systems in the wild.
Summary updated