Numerical modeling and analysis of machine learning based security framework for 6G communication system

6G, the next generation in communication systems, is implemented and presents both challenges and opportunities like have never been seen before. This work is centered on the combination of 6G technology and Artificial Intelligence (AI) attacking Security risks due to adversarial machine learning. Although the ML methods, particularly, deep learning can bring significant improvements to 6G applications such as mmWave beam prediction, they are also ‘significantly susceptible’ to adversarial attacks. Attacks on input data aim at making its predictions incorrect hence compromising the performance and integrity of 6G systems. These are categorized under adversarial machine learning attacks, and in this study, we use the Fast Gradient Sign Method (FGSM), to show how the efficiency of ML models in 6G for mmWave beam prediction can be compromised. In our use-case we observe that the Mean Squared Error (MSE) of the RF beamforming prediction model undergoes a sharp rise under attack. Towards this end, adversarial training is put forward as a defense mechanism. Finally, the experiment of retraining the model using adversarially modified input data shows that the model has better stability in the transformed regions, achieving a lower MSE than attacked cases, thus successfully prevent adversarial attacks. On the public dataset DeepMIMO, TensorFlow-based beam prediction models are trained and used to perform attacks. The findings presented here reveal that the model trained with adversarial data is more secure with minimum MSE compared to the model trained without defenses. The study therefore calls for inclusion of security features like adversarial training in the implementation of 6G communication systems to mitigate new forms of threat to AI interventions.