Reproduce “Adversarial Threats in Quantum Machine Learning: A Survey of Attacks and Defenses” on OpenResearch

About this paper

Archisman Ghosh, Satwik Kundu, Swaroop Ghosh

This paper comprehensively surveys adversarial threats and corresponding defense mechanisms in Quantum Machine Learning (QML), categorizing attacks by adversary access level and detailing how classical machine learning vulnerabilities are uniquely amplified in quantum systems. It consolidates research on model theft, side-channel attacks, and data poisoning in QML, while also reviewing defense strategies such as quantum logic locking, hardware-aware watermarking, and secure model partitioning across multiple backends.

Key insights

• QML systems, especially in cloud-based QMLaaS deployments, introduce unique attack surfaces due to NISQ hardware characteristics, hybrid architectures, and the opacity of quantum operations.
• Adversarial threats to QML span various levels of access, from black-box model extraction and side-channel attacks exploiting physical crosstalk, to white-box threats like circuit reverse engineering and pulse-level manipulation.
• Defense mechanisms often leverage quantum properties, such as hardware noise signatures for watermarking or circuit partitioning for secure distribution, alongside adaptations of classical security techniques like obfuscation and randomization.

Summary by alphaXiv.