Adversarial Training and Provable Robustness: A Tale of Two Objectives

Jiameng Fan, Wenchao Li

Abstract

We propose a principled framework that combines adversarial training and provable robustness verification for training certifiably robust neural networks. We formulate the training problem as a joint optimization problem with both empirical and provable robustness objectives and develop a novel gradient-descent technique that can eliminate bias in stochastic multi-gradients. We perform both theoretical analysis on the convergence of the proposed technique and experimental comparison with state-of-the-arts. Results on MNIST and CIFAR-10 show that our method can consistently match or outperform prior approaches for provable $l_\infty$ robustness. Notably, we achieve 6.60% verified test error on MNIST at $\epsilon = 0.3$, and 66.57% on CIFAR-10 with $\epsilon = 8/255$.

Introduction

Deep neural networks are notoriously vulnerable to adversarial attacks—small, intentional perturbations that flip a model's prediction. To combat this, researchers usually take one of two paths: empirical adversarial training or provable robustness verification. Adversarial training involves training on "hard" examples generated during training (e.g., using PGD), making the model empirically strong but offering no mathematical guarantees. Conversely, provable robustness uses verification techniques like Interval Bound Propagation (IBP) to calculate a mathematical upper bound on the worst-case loss, providing a "certificate" of safety but often resulting in lower overall accuracy and difficult tuning.

There is a significant "robustness gap" between what a model can actually withstand and what it can prove it can withstand. This paper introduces AdvIBP, a principled framework that bridges this gap by treating empirical and certified robustness as a joint optimization problem. By minimizing both the empirical "adversarial loss" and the mathematical "abstract loss" simultaneously, we can "squeeze" the true robust loss from both sides.

The AdvIBP Framework

The core of AdvIBP lies in combining Fast Gradient Sign Method (FGSM) adversarial training with IBP-based verification. The framework addresses the challenge of competing objectives through three main innovations.

Multi-Objective Optimization

Instead of a simple weighted sum of losses—which requires manual weight tuning and complex scheduling—AdvIBP treats training as a two-objective problem. Adversarial loss serves as a lower bound on the true robust loss, while abstract loss (from IBP) provides an upper bound. Minimizing both jointly is more effective than focusing on one and using the other as a regularizer.

Gradient Balancing

To handle potential conflicts between the adversarial gradient ($g_{adv}$) and the provable gradient ($g_{IBP}$), the algorithm employs an adaptive balancing mechanism. If the gradients point in the same general direction ($g_{adv}^T g_{IBP} > 0$), the algorithm follows their angular bisector. If they conflict, the algorithm uses the First-Order Stationary Condition (FOSC) to decide which objective to prioritize.

The AdvIBP gradient balancing mechanism for resolving objective conflicts.

As shown in the figure above, the FOSC value $c(x_{adv})$ allows the model to automatically sense the "strength" of current adversarial examples. If the attacks are weak ($c(x_{adv}) \le c_t$), the model prioritizes the mathematical verification bounds ($g_{IBP}$). If the adversarial objective is far from stationary ($c(x_{adv}) > c_t$), it prioritizes adversarial training ($g_{adv}$) to find a more robust region.

Eliminating Bias

A common issue in multi-gradient descent is that weights derived from stochastic gradients can be biased. AdvIBP utilizes moment estimates (similar to the logic in the Adam optimizer) to ensure the weights are independent of the current batch. This leads to more stable convergence and eliminates the need for the "ad-hoc" schedulers typically required in provable training.

Experimental Results

The framework was evaluated on MNIST and CIFAR-10 across various model sizes (DM-Small, DM-Medium, and DM-Large). AdvIBP consistently matched or outperformed previous state-of-the-art methods like CROWN-IBP and COLT in terms of verified test error—the percentage of test cases where the model is mathematically guaranteed to be correct within a certain perturbation range ($\epsilon$).

  • MNIST ($\epsilon = 0.3$): Achieved a verified test error of 6.60%.
  • CIFAR-10 ($\epsilon = 8/255$): Achieved a verified test error of 66.57%.

Despite the dual-objective approach, the training time remains comparable to standard IBP training, making it practical for larger models.

Conclusion and Future Work

AdvIBP demonstrates that empirical and provable robustness are not mutually exclusive but can be optimized jointly to achieve superior results. While the method currently relies on the computationally efficient but conservative IBP for upper bounds, future work could extend this framework to more precise methods like CROWN. Additionally, while the method scales well to CIFAR-10, scaling provable robustness to massive datasets like ImageNet remains an open challenge for the field.

Summary updated

Abstract

We propose a principled framework that combines adversarial training and provable robustness verification for training certifiably robust neural networks. We formulate the training problem as a joint optimization problem with both empirical and provable robustness objectives and develop a novel gradient-descent technique that can eliminate bias in stochastic multi-gradients. We perform both theoretical analysis on the convergence of the proposed technique and experimental comparison with state-of-the-arts. Results on MNIST and CIFAR-10 show that our method can consistently match or outperform prior approaches for provable l infinity robustness. Notably, we achieve 6.60% verified test error on MNIST at epsilon = 0.3, and 66.57% on CIFAR-10 with epsilon = 8/255.
BibTeX
@inproceedings{fan2021adversarial,
  author = {Jiameng Fan and Wenchao Li},
  title = {{Adversarial Training and Provable Robustness: A Tale of Two Objectives}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=AgAaT5BHYM},
}
Figure