QGAN: Quantum Generative Adversarial Network to Secure Deep Learning

Despite deep neural networks have demonstrated tremendous success in various commercial and DoD applications, they are susceptible to adversarial attacks with detrimental outcomes to the underlying applications. The generative adversarial network (GAN) provides a good way of defending against adversarial learning attacks, but it is faced with a practical challenge, as classical computers are not able to perform certain computations, e.g., solving non-convex optimization problems (as required by training the GAN), with the required level of speed to sustain applications. We propose to design and develop Quantum Generative Adversarial Networks (QGANs) by leveraging quantum computing techniques that will demonstrate robustness against adversarial attacks. Quantum computing can solve non-convex optimization problems in polynomial time. Hence, it can provide radically powerful capability to defend adversarial attacks on deep neural networks.