Millimeter-wave (mmWave) beam selection in MIMO systems presents significant challenges in dynamic environments due to computational constraints, data heterogeneity, and privacy concerns. In this paper, we propose a novel Maturing Federated Transfer Learning (MFTL) framework that integrates radar and image data to enhance beam prediction accuracy while ensuring user data privacy. The proposed approach utilizes ResNet-50 as a pre-trained model, fine-tuned locally at distributed Antenna Units (AUs) to adapt to diverse scenarios. To mitigate the effects of data heterogeneity, we evaluate multiple aggregation strategies, with FedMedian demonstrating superior robustness compared to FedAvg and weighted averaging. Our optimized MFTL configuration, utilizing a learning rate of 0.005, a batch size of 32, and five aggregation rounds, significantly improves model performance. Experimental results on the DeepSense 6G dataset indicate that our approach achieves a Top-1 accuracy of 57.58% and a Top-5 accuracy of${9 4. 7 \%}$, outperforming state-of-the-art methods. These findings highlight the effectiveness of federated learning, transfer learning, and robust aggregation techniques in improving beam selection accuracy for next-generation mmWave communication systems.

Federated learning (FL) has gained attention for enabling efficient distributed learning while maintaining data privacy. However, the data privacy constraint reduces the transparency in the agents’ model update making the learning process vulnerable to Byzantine attacks. In this paper, a mathematical proof is provided to show that when the traditional model-combining scheme is used, the model will eventually diverge to non-useful solutions in the presence of Byzantine agents independently from their number or their contributions. A low complexity norm-control based aggregation approach is also proposed and shown to converge to the optimal and sub-optimal solutions in the absence or presence of Byzantine nodes, respectively. Monte-Carlo simulations are also conducted to verify and validate the mathematical derivations and the efficiency of the proposed approach in protecting the FL model.

The abundance of social media data in the Arab world, specifically on Twitter, enabled companies and entities to exploit such rich and beneficial data that could be mined and used to extract important information, including sentiments and opinions of people towards a topic or a merchandise. However, with this plenitude comes the issue of producing models that are able to deliver consistent outcomes when tested within various contexts. Although model generalization has been thoroughly investigated in many fields, it has not been heavily investigated in the Arabic context. To address this gap, we investigate the generalization of models and data in Arabic with application to sentiment analysis, by performing a battery of experiments and building different models that are tested on five independent test sets to understand their performance when presented with unseen data. In doing so, we detail different techniques that improve the generalization of machine learning models in Arabic sentiment analysis, and share a large versatile dataset consisting of approximately 1.64M Arabic tweets and their corresponding sentiment to be used for future research. Our experiments concluded that the most consistent model is trained using a dataset labelled by a cascaded approach of two models, one that labels neutral tweets and another that identifies positive/negative tweets based on the Arabic emoji lexicon after class balancing. Both the BERT and the SVM models trained using the refined data achieve an average F-1 score of 0.62 and 0.60, and standard deviation of 0.06 and 0.04 respectively, when evaluated on five diverse test sets, outperforming other models by at least 17% relative gain in F-1. Based on our experiments, we share recommendations to improve model generalization for classification tasks.