Speaker
Description
Recent advances in quantum machine learning raise a basic security question: can malicious functionality be hidden inside a parameterized quantum classifier in a way that is both controllable by the attacker and difficult to recognize from the expected circuit structure? In classical machine learning, recent work on undetectable backdoors has shown that cryptographically structured triggers can be substantially more powerful than simple pattern-based triggers. In the quantum setting, however, the situation is less understood. Circuits are constrained by reversibility, measurements are limited, and the circuit itself may be inspected at least partially by a model owner, compiler, or auditor. We study a cryptographically grounded approach to this problem based on digital-signature verification. The central idea is to embed a verifier-like subroutine into a parameterized quantum classifier so that a signed input activates hidden malicious behavior, while unsigned inputs follow the ordinary classification pathway. We formulate the problem in a gray-box setting relevant to variational circuits, present three concrete construction paradigms, and compare them through the joint lens of trigger functionality, concealment strategy, and circuit-resource overhead. The analysis suggests that signature-based backdoors are a meaningful, previously underexplored primitive for quantum learning systems, while also highlighting a substantial gap between cryptographic elegance and practical embedding cost.
| Topic | Quantum variational circuits and their applications |
|---|