Adversarial Robust Quantum Reinforcement Learning: Defending Variational Policies Against Quantum-State Perturbations
Kainat Rizwan, Mir Wajid Ali
Quantum reinforcement learning (QRL) has shown potential significant potential for solving difficult sequence decision-making with the power of representation of quantum computing. However quantum states and their measurements are sensitive, and this implies that such systems are vulnerable to adversarial perturbation together with hardware noise. Avoiding variational quantum policy being misled by any but small, intelligible perturbations of quantum states, inputs or measurement outcomes can generate incorrect actions and unstable learning behaviour. This paper has provided an adversarially resilient quantum reinforcement learning (AR-QRL) architecture, which specifically aims to address these security and dependability challenges. It is a mixture of adversarial training, robust quantum states encoding and variance-sensitive policy optimization, which defends variational quantum policies against malicious quantumstate perturbation. Benchmarking quantum-control experiments have demonstrated improvements of 30%-40% in control fidelity to that of standard QRL, up to a 75% decreased reward variance, and a 40% acceleration of convergence under adversarial conditions.. These findings suggest the impact of adversarial robustness in improving effective and secure quantum reinforcement learning of noisy intermediate-scale quantum (NISQ) devices.
International Bhurban Conference on Applied Sciences and Technology (IBCAST)•Aug 1, 2026•0 citations