Computational hardness of robust classification proven under cryptographic assumptions

AI Impact Summary

This research demonstrates fundamental computational barriers to robust machine learning classification. The authors prove that certain classification tasks admit robust classifiers in theory but are computationally intractable to learn in practice—a gap that persists even when non-robust classifiers can be learned efficiently. Critically, they show this hardness applies to both small and large perturbation regimes under different cryptographic assumptions (one-way functions, learning parity with noise). The work establishes that robust classification hardness is deeply connected to cryptographic primitives, meaning any breakthrough in efficient robust learning would immediately yield new cryptographic constructions.