New Framework Certifies MLP Adversarial Robustness and Completeness.

Researchers have developed a new theoretical framework to address a fundamental challenge in AI safety: adversarial robustness. This framework redefines the problem as a lattice traversal, where each lattice element represents an interval containing an input point for a Multilayered Perceptron (MLP) classifier. The work distinguishes between "sound certification," which ensures an input can be perturbed within an interval without changing the MLP's prediction, and a novel concept of "complete certification," which guarantees the prediction *will* change if the input moves outside the interval. While sound certification is well-studied, complete certification is a new area of exploration. The proposed method uses iterative refine-and-verify steps with lattice traversal operators and formal MLP verifiers to ensure optimal sound and complete certifications. The study also reveals asymmetries in optimization problems, finding that minimum complete solutions are polynomially tractable, unlike sound certifications, and introduces logarithmic algorithms for symmetric intervals.