Testing of AI Models for Air-Interface Applications

Studenteropgave: Kandidatspeciale og HD afgangsprojekt

  • Filip Ivanovic
AI is a rapidly developing field of technology being implemented in many contexts, including in air interface communications. Alongside the development of AI comes the development of systems designed to test AI, ensuring its performance and understanding its characteristics. This project, done in collaboration with Keysight Technologies and using their xpl[AI]ned framework, stands as a study into the usability and relevance of a few different testing methods on AI models designed for the air interface communication space. This is done through a process of researching various different aspects of AI and testing methods and applying them to two AI models that are chosen to be representative of potential real world implementations. The two main testing methods applied are Monte Carlo dropout and the Fast Gradient Sign Method(FGSM) for testing adversarial robustness. Monte Carlo dropout is found to be useful in quantifying the efficiency of the construction of the models, but due to the nature of the context its quantification of certainty is not found to be useful. FGSM is found to be extremely useful with its capability to show if a model is vulnerable to adversarial perturbations, as well as generating adversarial data that can be analysed to further determine model characteristics. A few other methods and testing paths were also looked into to gain further clarity. The overall result of this investigation is a resounding success for the xpl[AI]ned concept as well as a greater understanding of the types of methods relevant when testing air interface AI models.
Udgivelsesdato30 maj 2023
Antal sider123
Ekstern samarbejdspartnerKeysight Technologies
Dr Alan Anderson alan.anderson@keysight.com
ID: 532287677