Quantum Machine Learning Fault Injection

This work stems from the composition of Quantum Computing, Machine Learning and Fault Injection and Reliability testing with the scope of understanding their interaction. Quantum Computing is still an emerging technology in constant evolution. The true extent of the advancements it will bring to humanity as a whole are numerous and possibly unpredictable. Machine Learning is comprised of all the techniques and attempts that aspire at infusing classical computation devices with what could be defined as "intelligence". Fault Injection and Reliability testing consists of analysis methods aimed at stressing electronic circuitry with the purpose of spotting out fault behaviours and patterns, eventually leading to the development of devices able to resist, detect and possibly correct such anomalies. After introducing one of the current models for quantum faults induced by external radiation on transmon-based devices, a metric for reliability measurement called Quantum Vulnerability Factor (QVF) is presented in depth. A purposefully-developed software simulation suite has been developed, in order to carry out circuit level fault injections, execute them in parallel on a number of different simulators or hardware backends and perform QVF based reliability analyses on quantum circuits. At the core of the work, the three fields are merged together in the fault reliability analysis of two modern quantum machine learning models, a Quantum Support Vector Machine and a Quantum Convolutional Neural Network. The aim is to study the interaction of these architectures with the radiation-induced transient quantum fault model. The Quantum Support Vector Machine QVF study allowed for the identification of the most critical faults in the circuit, demonstrating that, despite marginal differences, both qubits in the QSVM’s quantum subroutine have a similar level of reliability to faults. The highest criticality comes from phase shifts around the azimuthal angle of the Bloch sphere, whilst the polar angle shows a periodic response to faults which varies with respect to the distance from |+⟩. Furthermore, an analysis of the impact of a double fault has been investigated and compared to the one of the single fault. The QVF study on the four qubit circuit used in the Quanvolutional Neural Network gave a first description of its reliability pattern. Given the amplitude embedding strategy used for the inputs, it has been proven that an azimuthal angle based fault is devastating on the network’s output, a behaviour that highly different with respect to a polar angle based fault. To support this hypothesis also from the QNN’s accuracy side, a subset of fault positions has been studied more thoroughly, highlighting a response pattern which proved the previous assumption. This allowed for a final analysis at the dataset level, in order to spot out the input images which either maximised or minimized the network’s resilience to faults. Despite not being a thorough top down approach, it allowed to prove that there is a different resilience pattern depending on which qubit is affected by the fault, while at the same time showing that the depth of the fault in the circuit has a negligible impact on the final output. The study provided sufficient information to get a basic understanding of the circuits' characteristics, thus paving the way for further research in the future.