LO-Attack Defense Mechanism: Enhancing 6G Beam Prediction Model Security Against Complex Adversarial Attacks

As sixth-generation (6G) networks continue to evolve, the deployment of Machine Learning (ML) models in critical functions, such as millimeter-wave (mmWave) beam prediction, presents unique security challenges, particularly from adversarial attacks. Existing defenses against these attacks often have limitations, especially when gradient information is unavailable. This paper introduces the Lemur Optimizer for Adversarial Attacks (LO-Attack), a novel gradient-free approach inspired by swarm intelligence to enhance the robustness of 6G beam prediction models. By integrating LO-Attack into the adversarial training process, models become more resilient against both gradient-based and gradient-free attacks in dynamic, high-mobility environments. Experimental results from two 6G scenarios, encompassing indoor and outdoor settings, demonstrate that models trained with LOAttack achieve significant improvements in accuracy, resistance to adversarial perturbations, and computational efficiency. On average, adversarial training with LO-Attack enhances model robustness by 57.90%, outperforming traditional FGSM-based methods. These findings highlight LO-Attack's potential as an effective and scalable defense strategy for securing ML-driven applications in the complex 6G landscape.