TESTNAV Pareto-Guided Robustness Testing for Deep Learning

TESTNAV uses bi-objective optimization to navigate compositional perturbation spaces for model robustness testing. It prioritizes finding failures that maximize performance degradation while maintaining high input fidelity using metrics like SSIM and KID.