Abstract
Recent work has shown that it is possible to train deep neural networks that are provably robust to norm-bounded adversarial perturbations. Most of these methods are based on minimizing an upper bound on the worst-case loss over all possible adversarial perturbations. While these techniques show promise, they often result in difficult optimization procedures that remain hard to scale to larger networks. Through a comprehensive analysis, we show how a simple bounding technique, interval bound propagation (IBP), can be exploited to train large provably robust neural networks that beat the state-of-the-art in verified accuracy. While the upper bound computed by IBP can be quite weak for general networks, we demonstrate that an appropriate loss and clever hyper-parameter schedule allow the network to adapt such that the IBP bound is tight. This results in a fast and stable learning algorithm that outperforms more sophisticated methods and achieves state-of-the-art results on MNIST, CIFAR-10 and SVHN. It also allows us to train the largest model to be verified beyond vacuous bounds on a downscaled version of IMAGENET.
Introduction
Deep learning models are notoriously fragile to adversarial attacks—small, imperceptible perturbations that can cause a model to flip its prediction with high confidence. While empirical defenses like adversarial training provide some protection, they often fail against stronger, unforeseen attackers. The gold standard for security is formal verification: a mathematical proof that no perturbation within a certain range (usually an $\ell_\infty$ ball) can change the model's output.
However, existing verification methods face a significant scalability vs. tightness trade-off. Exact methods are too slow for training, while fast "incomplete" methods often produce bounds so loose they are "vacuous" for large networks. This research demonstrates that a simple and computationally efficient technique—Interval Bound Propagation (IBP)—can be scaled to provide tight, meaningful robustness guarantees for large-scale models.
Methods: Interval Bound Propagation (IBP)
The core of the proposed approach is Interval Bound Propagation (IBP). Instead of propagating a single point through the network, IBP propagates an "interval"—an axis-aligned bounding box—that encompasses all possible adversarial perturbations within a given $\ell_\infty$ ball.
While IBP is computationally efficient (costing roughly twice a standard forward pass), it is traditionally considered too loose for training. The authors introduce three key enhancements to overcome this:
- Elision of the Last Layer: Mathematically bypassing the relaxation of the final linear layer to significantly tighten the bounds on output logits.
- Specific Loss Function: Using a weighted combination of standard cross-entropy loss and a "worst-case" cross-entropy loss calculated using the IBP bounds.
- Training Curriculum: Gradually increasing the perturbation radius ($\epsilon$) and the weight of the robust loss. This allows the network to slowly adapt its weights so that the IBP bounds become tight over time.
Key Results
The researchers evaluated their approach on standard benchmarks including MNIST, CIFAR-10, SVHN, and a downscaled version of ImageNet. The method achieved state-of-the-art verified accuracy while maintaining high computational efficiency.
- State-of-the-Art Verified Accuracy: On MNIST ($\epsilon=0.3$), the verified error rate was reduced from 19.32% to 8.05%. On CIFAR-10 ($\epsilon=8/255$), it improved from 78.22% to 67.96%.
- ImageNet Milestone: The authors produced the first non-vacuous verified error rate (93.87%) for a model as large as ImageNet. While the error remains high, it represents a massive leap from previous methods where bounds were equivalent to random guessing.
- Efficiency: Training a small MNIST model took only 3.5 seconds per epoch, compared to 120 seconds for previous state-of-the-art methods. This speedup is what enables the jump to ImageNet-scale architectures like WideResNet-10-10.
Insights and Discussion
A surprising insight of this work is the "self-fulfilling prophecy" of IBP: while IBP bounds are loose for arbitrary networks, a network trained using IBP naturally learns to represent functions where those bounds are tight. The model effectively shapes its own latent space to fit the axis-aligned boxes of the IBP method.
Despite these gains, a "robustness-accuracy trade-off" remains; robust models typically perform slightly worse on unperturbed images than standard models. Future work will focus on further tightening these bounds and exploring how to reduce the performance gap between robust and standard models across more complex architectures.
Summary updated