Federated Learning-Based Cooperative Spectrum Sensing with Belief Accumulation

Federated Learning-Based Cooperative Spectrum Sensing with Belief Accumulation

Introduction

Spectrum sensing (SS) is a critical technology in cognitive radio networks that enables secondary users (SUs) to detect and utilize idle spectrum bands originally allocated to primary users (PUs). The key challenge in SS lies in accurately determining the presence or absence of PU signals, particularly in low signal-to-noise ratio (SNR) environments where traditional methods struggle. While deep learning (DL)-based SS techniques have shown superior performance, they are often limited by dataset constraints and susceptibility to environmental variations. Cooperative spectrum sensing (CSS) leverages spatial diversity among multiple SUs to enhance detection reliability. However, conventional CSS approaches face challenges in effectively fusing heterogeneous SU data while mitigating interference from poorly performing nodes.

This article presents a novel Federated Learning-based Cooperative Spectrum Sensing (FL-CSS) framework that integrates distributed learning with trust-based data fusion. The proposed method addresses two major challenges: (1) maintaining high detection probability in low-SNR conditions, and (2) minimizing interference from unreliable SUs during collaborative decision-making. The framework operates in two distinct phases: an offline training stage for model optimization and an online detection stage for real-time spectrum sensing. By incorporating a belief accumulation mechanism, FL-CSS dynamically adjusts SU contributions based on their historical performance, environmental conditions, and model reliability.

System Architecture and Methodology

Overview of FL-CSS Framework

The FL-CSS system consists of multiple multi-antenna SUs and a fusion center (FC) that coordinates collaborative sensing. The architecture separates model training from real-time detection to optimize computational efficiency. During offline training, SUs participate in federated learning to develop shared knowledge while preserving data privacy. The online detection phase implements a sophisticated fusion strategy that weights SU inputs according to their accumulated trust scores.

Key innovations include:

1. A log-based preprocessing technique that enhances signal correlation features in low-SNR conditions
2. A federated model update strategy that prevents global model degradation from unreliable local updates
3. A belief accumulation algorithm that dynamically adjusts SU weights based on multiple performance factors

Signal Model and Preprocessing

The system considers a scenario with a single-antenna PU and multiple SUs equipped with multiple antennas. Each SU captures time-domain signal samples that may contain either noise alone (under hypothesis H0) or a PU signal corrupted by noise (H1). The received signals undergo specialized preprocessing to improve feature extraction:

First, the SU normalizes the received signal matrix and computes correlation coefficients between antenna pairs. A logarithmic transformation then amplifies weak correlation patterns that are particularly valuable for low-SNR detection. This nonlinear processing helps distinguish between noise-dominated and signal-containing scenarios. The transformed correlation matrices are then separated into in-phase and quadrature components to preserve modulation characteristics, creating a two-channel input for subsequent neural network processing.

Local Model Design

Each SU employs a convolutional neural network (CNN) for local spectrum sensing. The network architecture balances complexity and performance with three convolutional layers for feature extraction followed by fully connected layers for classification. Key design considerations include:

• Small 3×3 convolutional kernels that preserve spatial relationships in the correlation matrices

• ReLU activation functions that introduce nonlinearity while maintaining gradient flow

• A two-dimensional max-pooling layer that reduces dimensionality while retaining important features

• A final softmax layer that produces probabilistic outputs for H0 and H1 hypotheses

The local training process minimizes a cross-entropy loss function that encourages the network to maximize the posterior probability of correct hypothesis classification. Importantly, the model learns to generate well-calibrated confidence scores that reflect true detection probabilities rather than just binary decisions.

Federated Learning for Model Optimization

Global Model Generation

The federated learning component addresses the challenge of maintaining robust detection performance across SUs operating in diverse environmental conditions. The FC aggregates local model parameters using a weighted averaging scheme where each SU’s contribution depends on its model’s classification performance.

The weighting mechanism uses area under the ROC curve (AUC) scores to quantify each local model’s detection capability. Models performing significantly better than random chance (AUC > 0.5) receive higher weights, while poorly performing models have reduced influence. This selective aggregation prevents the global model from being corrupted by SUs experiencing particularly adverse conditions.

Local Model Update Strategy

When receiving updated global models, SUs employ a specialized adaptation technique that balances global knowledge with local specialization. The FedProx algorithm modifies standard gradient descent by adding a regularization term that penalizes significant deviations from the global model parameters. This approach:

• Preserves valuable features learned collectively by all SUs

• Allows limited adaptation to local signal characteristics

• Prevents overfitting to potentially noisy local datasets

The regularization is applied selectively to the convolutional layers responsible for fundamental feature extraction, while the classification layers retain more flexibility to adapt to local conditions. This balanced update strategy enables SUs to benefit from collective learning while maintaining necessary specialization for their specific operating environments.

Trust-Based Data Fusion

Decision Statistics and Thresholding

The online detection phase combines probabilistic outputs from multiple SUs to make global spectrum occupancy decisions. Each SU’s local model produces a confidence score representing the likelihood of PU signal presence. The FC aggregates these scores using trust-based weights to form a composite decision statistic.

Threshold determination follows Neyman-Pearson criteria to maintain a constant false alarm rate. Using Monte Carlo simulations with historical H0 data, the system empirically determines the detection threshold that yields the desired false alarm probability. This approach avoids assumptions about underlying signal distributions that may not hold in practical scenarios.

Belief Accumulation Algorithm

The core innovation in the fusion strategy is the dynamic trust weighting mechanism that considers three key factors:

1. Model Contribution Weight (ρt): Reflects the SU’s importance in global model generation based on its AUC performance
2. Recent Decision Accuracy (ρd): Rewards consistent correct decisions and penalizes errors through adjustable gain factors
3. Output Stability (ρv): Reduces weights for SUs showing high decision variance, indicating unreliable or fluctuating conditions

The trust update equation combines these factors additively with carefully tuned parameters that maintain stability while responding meaningfully to performance changes. A sliding window tracks recent decision behavior to compute variance terms efficiently.

The algorithm automatically adapts to various challenging scenarios: • Static interference conditions where certain SUs experience persistent SNR degradation

• Dynamic environments with time-varying channel conditions

• Malicious attack scenarios where compromised SUs intentionally provide false information

Parameter selection follows analytical guidelines that ensure stable trust evolution while providing sufficient responsiveness to changing conditions. The balance between reward and penalty factors prevents excessive trust fluctuations that could degrade detection reliability.

Performance Evaluation

Local Sensing Performance

Comparative studies demonstrate that the proposed preprocessing and local CNN architecture significantly outperform conventional spectrum sensing techniques, particularly in low-SNR regimes. Key findings include:

• The log-based correlation matrix processing provides 3-5 dB SNR gain compared to raw correlation analysis

• The two-channel IQ decomposition improves detection probability by 15-20% compared to single-channel approaches

• The local model achieves superior performance compared to energy detection, covariance-based methods, and alternative DL architectures across all tested SNR conditions

Cooperative Sensing Advantages

System-level evaluations highlight the benefits of the federated learning framework and trust-based fusion:

1. Model Robustness: FedProx updates maintain stable performance even when some SUs train with limited or noisy local data
2. Weighted Aggregation: AUC-based model fusion improves global model quality compared to simple averaging
3. Collaborative Gain: Additional SUs provide consistent detection probability improvements when properly weighted

In comparative tests against conventional CSS methods (LRT, GLRT, SVM), FL-CSS shows particularly strong advantages in challenging low-SNR conditions while maintaining computational feasibility.

Resilience to Challenges

The belief accumulation mechanism proves especially valuable in adverse scenarios:

Static Interference: When one of three SUs experiences 2-4 dB SNR degradation, trust weighting preserves 80-90% of the interference-free detection performance, compared to 50-60% for unweighted fusion.

Dynamic Environments: Under time-varying interference, the algorithm successfully tracks SU reliability changes, adjusting weights to favor currently stable nodes. Detection probability remains within 15% of optimal compared to 40% variations in unweighted approaches.

Malicious Attacks: With one malicious SU providing inverted decisions, the system maintains correct global decisions in 85% of cases at -16 dB SNR. Even with two malicious nodes, the trust mechanism correctly identifies and mitigates the attackers in 70% of decisions, while unweighted fusion fails completely.

Implementation Considerations

Computational Complexity

The distributed nature of FL-CSS provides favorable computational characteristics:

• Local preprocessing requires O(PN + 2P²) operations for P antennas and N samples

• CNN inference complexity grows linearly with the number of convolutional channels and quadratically with antenna count

• FC operations primarily involve weighted averaging with minimal overhead

Communication costs are managed through: • Periodic rather than continuous model updates

• Selective parameter sharing (e.g., feature extractor only)

• Compact representation of confidence scores for online detection

Practical Deployment Factors

Successful implementation requires attention to: • Initial training data collection covering diverse operational scenarios

• Calibration of trust accumulation parameters for specific deployment environments

• Synchronization mechanisms for cooperative sensing intervals

• Security provisions for federated learning communications

The framework shows particular promise for applications including: • Dynamic spectrum access in 5G/6G networks

• Military and emergency communication systems

• IoT networks with intermittent spectrum opportunities

Conclusion

The FL-CSS framework represents a significant advancement in cooperative spectrum sensing by combining federated learning with intelligent data fusion. The system’s dual-phase operation separates computationally intensive model training from real-time detection while maintaining adaptability through continuous trust updates.

Key advantages include: • Enhanced low-SNR detection through specialized signal preprocessing

• Robust collaborative learning via performance-weighted model aggregation

• Adaptive interference mitigation via multidimensional trust evaluation

• Resilience to both environmental challenges and malicious behavior

Future research directions include extension to wideband sensing scenarios, incorporation of additional contextual information for trust computation, and development of more efficient model compression techniques for resource-constrained SUs. The principles demonstrated in FL-CSS may also find application in other distributed sensing problems requiring reliable collaboration among heterogeneous nodes.

https://doi.org/10.19734/j.issn.1001-3695.2024.06.0258

Was this helpful?

YesNo

0 / 0