Researchers reduce MLP safety to lattice traversal for robust AI
New sound and complete interval certifications guarantee adversarial robustness with provable guarantees.
A new paper presents a rigorous theoretical framework that reduces adversarial robustness in MLP classifiers to a lattice traversal problem. It defines sound intervals (prediction unchanged) and complete intervals (prediction changes outside the interval), with algorithms guaranteeing sound maximality and complete minimality via formal verifiers. The work uncovers an asymmetry: complete certification is solvable in polynomial oracle calls, while sound certification is strongly intractable. The approach is validated by the ParallelepipedoNN system.
- Reduces MLP adversarial robustness to a lattice traversal over axis-aligned intervals.
- Introduces complete certification (prediction changes outside interval) – previously unstudied in literature.
- Complete certification is polynomial-time tractable; sound certification is strongly intractable.
- Empirical validation via ParallelepipedoNN system on MLP classifiers.
Why It Matters
Provides provable guarantees for MLP safety, enabling verifiable AI in critical applications like autonomous driving or medical diagnosis.