TESTNAV Optimizes Robustness Testing for Deep Learning Models.

Deep learning models are known to be vulnerable to real-world input perturbations, especially when multiple corruptions, such as brightness shifts combined with motion blur, occur simultaneously. Testing for these "compositional" effects is challenging due to the combinatorial explosion of the perturbation space and the fact that many combinations result in unrealistic inputs with limited diagnostic value. To address these challenges, researchers present TESTNAV, a Pareto-guided robustness testing framework. TESTNAV is designed for efficiently exploring discrete, compositional perturbation spaces, particularly when the number of evaluable configurations is limited. It reframes robustness testing as a bi-objective optimization problem: the goal is to maximize performance degradation while simultaneously preserving the input's fidelity, which is measured using modality-specific metrics like SSIM for vision or BERT-F1 for language. The framework employs the NSGA-II algorithm to approximate the bi-objective Pareto front, effectively prioritizing severe yet realistic failures. Across four benchmarks spanning vision, natural language, and code generation, TESTNAV demonstrated significant efficiency gains. It recovered Pareto fronts up to 2.15 times faster than search-based baselines, utilizing only 35.8% to 89.3% of the discrete perturbation space defined by four perturbation dimensions with six levels each.