RepFair-QGAN: Alleviating Representation Bias in Quantum Generative Adversarial Networks Using Gradient Clipping

Kamil Sabbagh, Hadi Salloum, Yaroslav Kholodov

Keywords: Quantum Generative Adversarial Networks, GAN, Fairness

TL;DR: This study presents a novel application of Quantum Generative Adversarial Networks (QGANs) by introducing a group-wise gradient norm clipping technique to enhance representational fairness and mitigate mode collapse.

Abstract: This study introduces a novel application of Quantum Generative Adversarial Networks (QGANs) by incorporating a new fairness principle, \textit{representational fairness}, which improves equitable representation of various demographic groups in quantum-generated data. We propose a \textit{group-wise} gradient norm clipping technique that constrains the magnitude of discriminator updates for each demographic group, thereby promoting fair data generation. Furthermore, our approach mitigates the issue of mode collapse, which is inherent in both QGANs and classical GANs. Empirical evaluations confirm that this method enhances \textit{representational fairness} while maintaining high-quality sample generation.

Submission Number: 95