A machine learning approach to detecting crosstalk-induced fidelity degradation in quantum computing systems

Abstract Crosstalk, where operations on one qubit can perturb neighbours, has emerged as a significant security vulnerability in multi-tenant quantum computing environments. This phenomenon can be exploited as an attack vector, allowing malicious users to create side channels that leak proprietary information or to inject faults that degrade computational fidelity. We present a simulation-based investigation into the feasibility of using classical machine learning (ML) techniques to detect and quantify this fidelity degradation. Our proposed TripleBranchNet model was trained on a dataset derived from MNISQ, generated via a simulated backend based on IBM’s 127-qubit Brisbane device, and injected with various noise profiles, including crosstalk. TripleBranchNet integrates three specialised branches: a convolutional neural network (CNN) branch to analyse the off-diagonal elements of the density matrix, a multilayer perceptron (MLP) branch for its diagonal elements, and a second MLP branch for classical circuit descriptors. Trained on 80% of the data, the model achieved a mean absolute error of 0.0055 and an R 2 score of 0.5131 on unseen test data. The results indicate that crosstalk leaves a detectable imprint in the density matrices that our classical AI model can learn. This supports the feasibility of software-level monitoring for crosstalk-induced fidelity degradation without any hardware-level modifications. However, a significant portion of the variance (49%) remains unexplained. Our work thus offers a preliminary, proof-of-concept towards AI-assisted safeguards for growing quantum multi-tenancy, but real-device validation is still essential as a practical step forward.