HQMSGAN: Enhancing Adversarial Robustness Using Hybrid Quantum-Classical Multi-Scale Generative Adversarial Network

The growing vulnerability of computer vision models to adversarial attacks has raised great concerns about the reliability of generative models in security-sensitive applications. Quantum generative models show strong potential for learning complex data distributions. However, the robustness of quantum generative models to adversarial perturbations and the effectiveness of adversarial training using quantum-classical generative architectures remain less explored. In this work, we evaluate quantum and classical generative network-inspired adversarial training to analyse whether quantum state-space representations can enhance adversarial robustness compared with their classical counterparts. We propose HQMSGAN, a novel hybrid quantum-classical generative adversarial network with multiscaling discriminators that denoise the adversarial attacks through adversarial training. During adversarial training, the proposed model performs quantum projections to learn rich nonlinear feature representations using quantum state encoding and performs feature coupling using quantum entanglement in the quantum state space. Thus, it learns the probability distribution of both the clean and adversarially perturbed data and tries to generate pure samples by reducing the loss over iterations in the quantum realm. The experiment is carried out by inducing four different adversarial attacks and defending them by adversarial training using the resources provided by AWS Braket's quantum simulators in the Noisy Intermediate Scale Quantum (NISQ) era. The results provide favorable performance on metrics such as Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), Fréchet Inception Distance (FID) score, Learned Perceptual Image Patch Similarity (LPIPS), and classification accuracy after performing adversarial training.