Description
Cloud quantum computing is crucial for developing innovative quantum circuits
but also presents a potential threat model when using quantum computing
as a cloud service. The cloud quantum provider may act as an adversary
by advertising the use of quantum computers to clients while using cheaper
simulators to return results. This thesis demonstrates that it is possible to distinguish
whether a quantum circuit has been executed by a simulator (possibly
with noise) or a quantum computer with high accuracy. Three machine
learning techniques are being utilized: a support vector machine, a feedforward
neural net, and a convolutional neural net. Additionally, an adversarial
attack has been used to test the resilience of the feedforward neural net.
The test results showed that the model is vulnerable to adversarial samples,
reducing the model’s accuracy. The machine learning techniques learn the
noise fingerprint, and the results of this work are consistent with a previous
paper, indicating that it is feasible to differentiate between various quantum
computers by their noise fingerprint. All three machine learning techniques
used in this thesis can distinguish between quantum computers and simulators
with an accuracy of over 99.9%. Even for previously unseen quantum
computers and simulators, these models classify with at least 80% accuracy
(the best model has at least 92% accuracy).
|
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching