Investigation of the Security of ML-models in IoT Networks from Adversarial Attacks
Denis Parfenov, Lyubov Grishina, Leonid V. Legashev, Artur Yu. Zhigalov, Anton Parfenov · 2023
Due to the growth of digital intelligent technologies and the introduction of applied machine learning models in various industries, the demand for their protection has increased. The security of computing systems from adversarial attacks is an area of research of cybersecurity algorithms, which have become particularly relevant at the moment. The purpose of this work is to study the effectiveness of targeted adversarial attacks on machine learning models based on tabular data, which are based on gradient methods for optimizing the loss function. Within the framework of this study, an approach to conducting an adversarial attack using the Low ProFool algorithm is considered, and an approach to the use of generative-adversarial networks for generating synthetic adversarial samples based on substitution of real values of the output feature is proposed. The basic machine learning models are based on a set of data generated on the DeepMIMO platform according to the ray tracing scenario in open space, in order to identify the presence of end devices with the base station. The results obtained can be used in the development of secure models for the convergence of artificial intelligence and the Internet of Things, the future direction of research includes the development of methods to counter the considered adversarial attacks.