ROBUSTNESS EVALUATION OF QUANTUM MACHINE LEARNING FOR NETWORK SECURITY: A SYSTEMATIC MAP AND CRITICAL APPRAISAL

There is not much information about detecting robustness of quantum machine learning detectors even though the idea of using the latter for network attack detection has already been made. Articles from Scopus and IEEE Xplore databases were retrieved in English for research published after 2017 and searched in the last week of September of 2026. A two-level screening plan was applied with the first level mapping 363 abstract records and the second level performing a check of robustness claims made in abstracts against their corresponding full text articles and a final level assessment against a 14-item rubric. Overall, 204 reports were retrieved of which only 76 were selected for the analysis, and finally 33 studies were selected of which three were identified via other methods. Robustness was evaluated mainly as simulated quantum noise (17 studies) and as distribution shift within single benchmarks (10). Attacks on the detector were tested in nine studies; poisoning in one; and attacks such as backdoors or certified robustness were not found at all. Gradient masking was only partly implemented without an in-depth check; also, not a single attack achieved the level of being realizable and optimal for traffic simultaneously, and not a single noise or hardware study was able to perturbed a classical baseline in the same way in that exact way. Significance testing was absent from 25 studies, and 25 relied on simulation alone. When an abstract claimed to be robust without naming a test, full texts of 18 out of 24 articles were found not to have any test reported, and none of the articles had their conclusions strictly supported by the evidence when judged under an explicit adjudication rule. The method used is not free from errors since it is based on AI-assisted screening, driven retrieval mainly through access, and also appraiser's poor agreement on the claims item. The argument of QML that makes network detectors more robust is not supported at the moment; a minimum evaluation and reporting checklist as well as a research roadmap are proposed to help.