Bypasses expensive formal verification solvers by designing neural networks that are 'verifiable by design' using the fast trivial Lipschitz bound.
March 31, 2026
Original Paper
Lipschitz verification of neural networks through training
arXiv · 2603.28113
The Takeaway
Instead of using computationally heavy SDP or MIP verifiers after training, this paper forces the simple layer-wise product bound to be tight during training. It demonstrates Lipschitz bounds orders of magnitude lower than previous work, enabling real-time safety and robustness guarantees.
From the abstract
The global Lipschitz constant of a neural network governs both adversarial robustness and generalization.Conventional approaches to ``certified training" typically follow a train-then-verify paradigm: they train a network and then attempt to bound its Lipschitz constant.Because the efficient ``trivial bound" (the product of the layerwise Lipschitz constants) is exponentially loose for arbitrary networks, these approaches must rely on computationally expensive techniques such as semidefinite prog