Quantum Generative Adversarial Networks for Image Generation and Adversarial Image Detection

A. Shyam Sundhar, J. Angelin Jeba, Bhopendra Singh, S. J. Vimal Aravintha, Sureshkumar Somayajula, Ahmed J. Obaid Department of Information and Communication, Tampere University, Tampere, Pirkanmaa, Finland. Department of Electronics and Communication Engineering, S.A. Engineering College, Chennai, Tamil Nadu, India. Department of Engineering, Amity University Dubai, Dubai, United Arab Emirates. Department of Data Science, Indiana University Bloomington, Bloomington, Indiana, United States of America. Department of Computer Science and Technology, Sunlife Canada Financials, Toronto, Ontario, Canada. Department of Computer Science, Faculty of Computer Science and Mathematics, University of Kufa, Kufa, Najaf Governorate, Iraq.

Abstract: Adversarial attacks pose a major challenge to classical deep learning systems, often leading to misclassification and reduced model reliability. To address this limitation, researchers propose a hybrid Quantum Generative Adversarial Network (QGAN) that integrates quantum feature extraction with classical adversarial learning. The model employs a four-qubit parameterized quantum generator to produce representative image features, which are then upsampled to 28×28 images. At the same time, a classical neural network discriminator distinguishes real samples from generated outputs and from adversarial perturbations. This hybrid design leverages the unique representational capacity of quantum circuits and the learning efficiency of classical networks, thereby achieving greater robustness and adaptability. The system is trained and evaluated on the MNIST dataset, a standard benchmark for image classification, allowing validation of generative fidelity and adversarial resistance. However, challenges arise due to the noise sensitivity of quantum circuits, limitations of small-qubit generators, and the need to balance convergence in adversarial training. Despite these constraints, the QGAN successfully produces visually realistic images while enabling the discriminator to identify subtle adversarial manipulations with higher precision. This work highlights the potential of hybrid quantum-classical models as resilient frameworks against adversarial attacks. It offers a pathway toward secure, quantum-enhanced machine learning pipelines for future large-scale applications.

---

Keywords: Quantum Generative Adversarial Network (QGAN); Image Generation; Adversarial Image Detection; Quantum Machine Learning; Classical Discriminator; MNIST Dataset; Quantum Computing.

---

Received on: 12/08/2025Revised on: 23/10/2025Accepted on: 16/11/2025Published on: 03/03/2026

---

---