Dataset schema for classifier training.

This paper proposes a hybrid ensemble classifier with denoising autoencoder (ECDAE) framework to address reliability and robustness challenges in cooperative spectrum sensing (CSS) for cognitive radio networks (CRNs). The proposed framework first employs an ensemble classifier (EC) to dynamically reconfigure the sensing time, optimizing performance while minimizing cost. The EC accurately estimates the sensing samples based on target detection probabilities, false alarm rates, and channel conditions. Subsequently, a denoising autoencoder (DAE) eliminates soft-combined energies from false-sensing users (FSUs) before soft fusion. The results show that EC surpasses other methods, including random forest (RFC), neural networks (NN), decision trees (DT), k-nearest neighbors (KNN), Gaussian naive Bayes (GNB), RUSBoost and XGBoost, achieving an F1 score of 99.23%, an accuracy of 99.78%, and a Matthews correlation coefficient (MCC) of 99.6%. Furthermore, optimized sensing time through EC is combined with DAE reconstruction delivers superior sensing performance at the fusion center (FC) producing low error probabilities compared to traditional schemes such as identical gain combination (IGS), highest gain combination (HGS), particle swarm optimization (PSO), differential evolution-based machine learning (DE-ML) and convolutional neural networks (CNN). On average, the ECDAE framework achieves a 99.4% and 98.1% reduction in error probability compared to traditional schemes (IGS, HGS) and a 92.7% to 97.2% reduction compared to advanced methods (PSO, DE-ML, CNN), across all tested SNR conditions and false-sensing attack scenarios. The framework maintains robustness across four distinct false-sensing scenarios: (1) no false sensing (NFS) reporting an low-energy signals, (2) yes false sensing (YFS) reporting consistently an always high-energy signals, (3) opposite false sensing (OFS) reporting an always invert decisions to the true energy states, and (4) yes/no false sensing (YNFS) where FSU randomly alternates between YFS and NFS - ensuring minimal error probabilities in global decisions.

本论文提出一种搭载降噪自编码器（denoising autoencoder, DAE）的混合集成分类器（ensemble classifier with denoising autoencoder, ECDAE）框架，以解决认知无线电网络（cognitive radio networks, CRNs）中协作频谱感知（cooperative spectrum sensing, CSS）面临的可靠性与鲁棒性挑战。所提框架首先采用集成分类器（ensemble classifier, EC）动态重构感知时长，在优化感知性能的同时最小化感知成本；该集成分类器可基于目标检测概率、误警率与信道状态，对感知样本进行精准估计。随后，降噪自编码器（denoising autoencoder, DAE）会在软融合前滤除来自误感知用户（false-sensing users, FSUs）的软合并能量。实验结果表明，EC的性能优于随机森林（random forest, RFC）、神经网络（neural networks, NN）、决策树（decision trees, DT）、k近邻（k-nearest neighbors, KNN）、高斯朴素贝叶斯（Gaussian naive Bayes, GNB）、RUSBoost与XGBoost等其他方法，其F1分数可达99.23%，准确率为99.78%，马修斯相关系数（Matthews correlation coefficient, MCC）达99.6%。进一步地，通过集成分类器优化后的感知时长与降噪自编码器的重构结果相结合，可在融合中心（fusion center, FC）实现更优异的感知性能，相较于等增益合并（identical gain combination, IGS）、最大增益合并（highest gain combination, HGS）、粒子群优化（particle swarm optimization, PSO）、基于差分进化的机器学习（differential evolution-based machine learning, DE-ML）与卷积神经网络（convolutional neural networks, CNN）等传统方案，其全局决策的误判概率更低。在所有测试的信噪比（Signal-to-Noise Ratio, SNR）条件与误感知攻击场景下，相较于传统方案IGS与HGS，ECDAE框架的误判概率平均可降低99.4%与98.1%；相较于PSO、DE-ML、CNN等先进方法，其误判概率平均可降低92.7%至97.2%。该框架在四类典型误感知场景下均能保持鲁棒性：（1）无误感知（no false sensing, NFS）：上报低能量信号；（2）持续误感知（yes false sensing, YFS）：始终上报高能量信号；（3）反向误感知（opposite false sensing, OFS）：始终输出与真实能量状态相反的决策；（4）随机交替误感知（yes/no false sensing, YNFS）：误感知用户在YFS与NFS模式间随机切换，从而确保全局决策的误判概率维持在极低水平。