Optimized Double‐Stream Kolmogorov–Arnold Transformer Network for Secure 6G Millimeter‐Wave Beam Prediction

Accurate and efficient beam prediction is essential for reliable millimeter‐wave communication in emerging sixth‐generation (6G) networks. However, high‐dimensional channel data, noise sensitivity, and computational complexity limit the real‐time applicability of existing deep learning models. This paper proposes a lightweight and robust beam prediction framework that integrates curve‐based signal preprocessing, hybrid feature selection, and a dual‐stream Kolmogorov–Arnold transformer‐convolutional architecture optimized through Cleaner Fish Optimization. The preprocessing stage enhances signal stability, whereas the hybrid optimization strategy selects compact and informative features to reduce redundancy. The proposed dual‐stream architecture improves nonlinear representation efficiency without incurring high computational overhead. Experimental results demonstrate superior prediction accuracy, reduced inference time, and lower computational cost compared with conventional convolutional neural network, recurrent, and transformer‐based models. The framework achieves near‐perfect accuracy while maintaining real‐time suitability, making it a promising solution for scalable and trustworthy 6G millimeter‐wave beam management.