@ericyocam6Shipped: Classical vs. Quantum AI — Adversarial Robustness Comparison
An open-source AI security research proof of concept comparing the adversarial robustness of classical Convolutional Neural Networks (CNNs) and hybrid quantum-classical neural networks (HNNs).
Using the MNIST handwritten digit dataset, the framework trains and evaluates both architectures before and after individual and compounded white-box adversarial attacks.
The hybrid architecture integrates a classical CNN with a simulated parameterized quantum circuit containing rotation and entanglement operations.
The experiments examine how adversarial perturbations affect classification accuracy, prediction confidence, and model behavior across different architectures.
Results include model training curves, classification outcomes, adversarial examples, and comparisons of pre-attack and post-attack performance.
The primary research question is whether hybrid quantum-classical architectures demonstrate different robustness characteristics from conventional neural networks under controlled attack conditions.
The experiments are educational research demonstrations and do not establish a general quantum advantage or universal resistance to adversarial attacks.
Comments2
Running the same attacks on both versions, first one at a time and then stacked together, gives people a fair side-by-side instead of a vague claim that quantum is safer. Students who watch the same handwritten digits fool both systems will learn more from that than from any lecture on the topic.
Thanks, Ryan! That's exactly the educational value I wanted to demonstrate. By applying the same individual and compounded adversarial attacks to both classical and hybrid quantum-classical models, students can directly compare their vulnerabilities rather than assume one architecture is inherently more secure. Seeing how small input perturbations can affect classification also makes adversarial machine learning much more tangible. The goal is to understand where each architecture succeeds or fails, without overstating any quantum advantage. Appreciate your perspective!
2 comments
Running the same attacks on both versions, first one at a time and then stacked together, gives people a fair side-by-side instead of a vague claim that quantum is safer. Students who watch the same handwritten digits fool both systems will learn more from that than from any lecture on the topic.
Thanks, Ryan! That's exactly the educational value I wanted to demonstrate. By applying the same individual and compounded adversarial attacks to both classical and hybrid quantum-classical models, students can directly compare their vulnerabilities rather than assume one architecture is inherently more secure. Seeing how small input perturbations can affect classification also makes adversarial machine learning much more tangible. The goal is to understand where each architecture succeeds or fails, without overstating any quantum advantage. Appreciate your perspective!