Title: Reassessing Evaluation Frameworks for Certified Training in IBP-Based Methods

Abstract:

While deep neural networks excel in numerous supervised learning domains, they remain susceptible to adversarial perturbations. Although neural network verification offers mathematically rigorous robustness guarantees, it comes with significant computational overhead. Certified training addresses this challenge by optimizing for verifiable robustness during the training phase; however, this process typically creates a trade-off between natural accuracy and certified accuracy, governed by method-specific hyperparameters. Since these metrics are inherently at odds, the standard practice of reporting results from a single configuration is flawed: it can distort conclusions regarding overall performance and hinders unbiased evaluations of the current state of the art.

To resolve this, we propose evaluating certified training methods through Pareto front comparisons that map the natural-certified accuracy trade-off. By employing efficient automated multi-objective hyperparameter optimization, we identify a set of Pareto-optimal configurations for each method, ensuring fair and method-agnostic comparisons. This strategy frequently reveals that previously reported configurations were substantially undertuned, leading to improved performance and the establishment of a new state of the art. Utilizing these Pareto fronts, we deliver the first comprehensive multi-objective comparison of certified training approaches. Our findings indicate that prior advancements are less significant than previously thought and highlight previously unreported performance complementarities among different methods.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC