Exact and Asymptotically Complete Robust Verifications of Neural Networks via Quantum Optimization

Researchers have developed quantum optimization models for robust verification of deep neural networks against adversarial attacks. The approach provides exact verification for ReLU networks and asymptotically complete verification for networks with general activation functions like sigmoid and tanh.

• →New quantum optimization models can exactly verify robustness of neural networks with piecewise-linear activations like ReLU.
• →For general activation functions, the method provides scalable over-approximations that become increasingly accurate with refinement.
• →Quantum Benders Decomposition with interval arithmetic accelerates the solving process for robustness verification.
• →Certificate-transfer bounds enable relating robustness guarantees between pruned and original neural network models.
• →Experiments demonstrate high certification accuracy, showing quantum computing's potential for neural network security verification.