Rethinking Evaluation Paradigms in IBP-based Certified Training
Researchers have proposed a new method for evaluating certified training techniques in deep neural networks. Current practices often report a single configuration, which can be misleading due to the inherent trade-off between natural and certified accuracy. The new approach uses Pareto front comparisons to assess multiple configurations, enabling fairer and more comprehensive evaluations. This method has revealed that many previously reported configurations were undertuned, leading to superior performance and establishing a new state of the art in verifiable robustness. AI
IMPACT Establishes a more rigorous framework for evaluating AI model robustness, potentially leading to more reliable and secure AI systems.