Security of Artificial Intelligence Systems in 5G Networks

The adoption of the 5G technology in telecommunications and rapid growth in numbers, variety and density of connected devices raises serious security concerns. The increasing amount of network traffic and complexity of cyberattacks require the implementation of AI systems in network security. Such applications need thorough assessment of the resistance to data input manipulation by sophisticated adversaries. Conventional performance evaluation techniques for learning systems assume that datasets for training and validation of the model belong to the same benign environment and sufficiently represent the addressed phenomena. This assumption is violated in adversarial settings, where deployed systems are susceptible to a carefully crafted deceptive input resulting in performance degradation.
The dissertation project investigates the robustness of machine learning models to malicious samples given the specific constraints of the attacker's capabilities in 5G networks. Better understanding of the potential vulnerabilities of machine learning and optimal attack strategies are essential for the design of effective countermeasures.