Interval Certifications for Multilayered Perceptrons via Lattice Traversal

Short summary

This paper reduces the adversarial robustness problem for multilayered perceptrons to a lattice traversal problem, introducing both sound certifications (input can be perturbed within an interval without changing prediction) and complete certifications (prediction changes outside the interval). The authors develop lattice traversal operators in a refine-and-verify scheme using formal MLP verifiers, discovering asymmetries: complete certification minimization is polynomial while sound certification is strongly intractable. The ParallelepipedoNN system demonstrates the approach empirically.

• •Adversarial robustness for MLPs reduced to lattice traversal with sound and complete interval certifications
• •Complete certification minimization is polynomial; sound certification proves strongly intractable
• •ParallelepipedoNN system provides empirical validation of the framework

Generated with AI, which can make mistakes.