Adversarial Robustness in Fake Quantum Simulators
This paper investigates the performance scalability and adversarial robustness of Quantum Machine Learning (QML) models deployed on noise-model-based fake simulators. We conduct a dual-phased study, first benchmarking the computational throughput of Qiskit's Aer simulation engine across varying hardware architectures, and second, evaluating the effectiveness of Projected Gradient Descent (PGD) attacks and adversarial retraining strategies under realistic noise conditions. Our results quantify the runtime tradeoffs and scaling behavior for medium-scale simulations (projected up to 8 qubits) and demonstrate that high adversarial-to-benign retraining ratios (50/50) are essential for achieving practical model robustness for 4-qubit classifiers under realistic noise conditions.