Federated Learning (FL) has emerged as a promising paradigm for Network Intrusion Detection Systems (NIDS) by enabling distributed model training without sharing raw data. However, FL remains highly vulnerable to poisoning attacks, where malicious clients manipulate local updates to degrade global model performance. This study introduces WeiDetect, a Weibull distribution-based defense framework designed to identify and mitigate poisoning behaviors in FL-based NIDS environments. The proposed method models the statistical distribution of client-level gradient deviations using the Weibull probability distribution. By characterizing update irregularities through shape and scale parameters, WeiDetect distinguishes benign clients from adversarial participants without requiring labeled attack data. The framework integrates a robust aggregation mechanism that dynamically assigns weights to client updates based on their likelihood of conformity to learned Weibull behavior patterns. Experimental evaluation conducted on benchmark intrusion detection datasets demonstrates that WeiDetect consistently improves robustness under non-IID settings and various poisoning intensities. The model achieves higher detection accuracy and improved F1-score compared to conventional aggregation strategies such as FedAvg, trimmed mean, and median-based FL defenses. Additionally, WeiDetect maintains stable convergence behavior even under high adversarial participation rates, indicating strong resilience to model manipulation attacks. Experimental evaluations on benchmark intrusion datasets demonstrate that WeiDetect achieves 93.0% accuracy, 93.0% precision, 92.8% recall, and 92.9% F1-score under 30% poisoning conditions, significantly outperforming FedAvg, Trimmed Mean, and clustering-based defenses. Additionally, WeiDetect maintains a robustness index above 0.93, indicating strong stability under adversarial participation and non-IID data distributions.
The advent of 5th generation (5G) wireless networks and Internet of Things (IoT) has led to the development of the need for privacy-aware and precise traffic predictions in network slicing scenarios. Centralized machine learning models usually encounter various obstacles due to data privacy, high communication costs, and scalability concerns. In order to mitigate such problems, this paper presents a new method of Secure Uncertainty-Aware Federated Learning (Secure UA-FL), which utilizes Bayesian LSTM models in combination with Monte Carlo Dropout (MCD) technique to predict traffic uncertainty in distributed edge nodes. The model is based on the uncertainty-aware aggregation, which adjusts client weights adaptively during the training process, and uses Krum-based Byzantine defense approach for improved performance and security. The proposed model is tested on synthesized 5G traffic datasets corresponding to eMBB, URLLC, and mMTC 5G network slices. The Secure UA-FL framework provides a scalable, secure, and privacy-aware solution for traffic prediction in dynamic 5G network slicing scenarios.
Cryptocurrency price forecasting is inherently challenging due to pronounced volatility, non-stationary behaviour, and strong sensitivity to market sentiment. In this work, a comparative analysis is conducted between the recurrent Long Short-Term Memory (LSTM) network and the attention-based Time Series Transformer (TST) for predicting Bitcoin prices. Daily Bitcoin price data are modelled using three feature configurations: (i) raw OHLCV variables, (ii) OHLCV combined with technical indicators, and (iii) an extended feature set incorporating public interest measured through Google Trends for the term Bitcoin. To examine temporal sensitivity, the TST model is evaluated across multiple lookback windows of 30, 60, and 90 days. Experimental results reveal that LSTM with OHLCV features achieves the lowest absolute forecasting error (MAE: 3,958; RMSE: 5,621), while feature enrichment degrades performance in both models. The TST demonstrates greater robustness across lookback window configurations, with its MAE varying within a narrower range (10,987–12,406) compared to LSTM under feature enrichment. These findings contribute nuanced insights into the conditions under which attention-based architectures offer practical advantages over recurrent models for cryptocurrency price prediction.
Non-orthogonal Multiple Access (NOMA) enhances system radio capacity and offers superior spectral efficiency. However, the intricacy of Successive Interference Cancellation (SIC) poses a significant challenge for NOMA implementation. To address this, dividing users into orthogonal frequency groups and employing hybrid NOMA within each group is recommended. Our simulation results for single-carrier NOMA (SC-NOMA) sum rate analysis demonstrate that increasing the number of users beyond a certain threshold leads to a decline in network sum throughput. Consequently, it’s not feasible to indefinitely increase the number of users per carrier. Furthermore, we investigated user pairing strategies within the hybrid NOMA framework. Our simulation results for Near-Near (N-N), far-far (F-F), and near-far (N-F) user pairing schemes demonstrate that user pairing can significantly reduce computational costs. To further evaluate network sum rate performance, we compared these pairing schemes using both SC-NOMA and Time Division Multiple Access (TDMA) over a Nakagami-m fading channel. Our findings show that pairing the nearest user with the farthest user dramatically increases the total sum rate. While NOMA still outperforms TDMA with respect to sum rate when using N-N and F-F pairings, the improvement is less pronounced. SC-NOMA, on the other hand, underperforms TDMA. At low Signal-to-Noise Ratios (SNR), N-N, F-F pairing yields a higher sum rate than N-F pairing. However, this trend reverses at high SNR, where N-F pairing offers a superior sum rate. Consequently, N-F pairing can also lead to increased sum rates at high SNR.
The multi-source localization in sensor array networks is a major challenge for radar and sonar. Recent advances in array signal processing techniques and hardware technologies have resulted in a significant increase in the accuracy of angle of arrival or time difference of arrival estimation. So, the angle of arrival or time difference of arrival is widely used as the fundamental features in source localization. In this paper, we address the problem of locating multiple sources based on the angle of arrival and time difference of arrival obtained from small sensor arrays. We propose a new likelihood function for multi-source localization under the assumption that the measurements follow a normal distribution. And we propose a method to simultaneously estimate the source positions using the genetic algorithm. Compared the existing methods, the effectiveness of the proposed method is demonstrated through simulation experiments.
