A Comparative Study of Machine Learning Techniques for Detecting Spectrum Sensing Attacks

Abstract

Cooperative Spectrum Sensing (CSS) facilitates effective spectrum sharing in CRNs through the collaborative detection of PU activity by SUs[1]. Even though cooperation enhances detection robustness against fading and interference, it simultaneously poses threats from SSDF attacks in which colluding SUs manipulate sensing data to affect the global decision made at the fusion center. Many existing approaches to detecting SSDF have claimed high accuracy, but their assumptions about wireless channels may not reflect reality. The paper describes a framework for SSDF detection through machine learning on datasets generated by simulations. The sensing behaviors are modeled according to different SNR levels, and adversarial manipulation is simulated using probabilistic rules to generate a heterogeneous dataset. Statistical features based on energy detection are extracted and fed into three classifiers: Support Vector Machine, Logistic Regression, and Random Forest. The results indicate that Random Forest consistently outperforms other analysed classifiers for different SNR and attack conditions Thus, tree-based algorithms can efficiently model complex non-linear relationships in feature space. The good performance of Random Forest indicates that the tree-based ensemble approaches are well suited to detect the non-linear relations that are typical in the feature space.

Introduction

The text presents a machine-learning-based approach for detecting Spectrum Sensing Data Falsification (SSDF) attacks in Cognitive Radio Networks (CRNs). The main objective is to improve the reliability and security of Cooperative Spectrum Sensing (CSS) under realistic wireless conditions.

• Cognitive Radio Networks allow Secondary Users (SUs) to opportunistically use unused spectrum bands when Primary Users (PUs) are not transmitting. SUs perform spectrum sensing to determine whether a PU is present.
• Cooperative Spectrum Sensing (CSS) improves detection reliability by combining sensing reports from multiple SUs. However, it assumes that users provide trustworthy information.
• SSDF attacks violate this assumption. Malicious SUs intentionally modify their sensing reports, which can cause:
  • increased false alarms and reduced spectrum availability, or
  • missed PU detections and harmful interference.
• Detecting malicious users is difficult because real wireless environments contain noise, fading, SNR variations, channel variability, uncertain observations, unexpected attacker behavior, and labeling errors.
• Traditional approaches such as statistical filtering and reputation-based methods may not be sufficiently robust in dynamic environments.
• The research therefore proposes a machine-learning-based SSDF detection framework using realistic simulated data rather than idealized datasets.

Research objective

The study evaluates the robustness of CSS against malicious users using a large-scale simulation involving:

• 200 Secondary Users
• 150 sensing iterations
• SNR ranging from −20 to +10 dB
• 5%–40% malicious-user attack ratios
• Approximately 1.05 million samples

Energy detection is used to generate sensing observations, and statistical features describing user behavior are extracted to identify malicious users performing decision-flipping attacks.

System model

Each SU determines whether a Primary User is present using two hypotheses:

• H?: PU is absent; the received signal consists primarily of noise.
• H?: PU is present; the received signal contains the PU signal plus noise.

The energy detection method calculates the received signal energy over a sensing period. The noise level is controlled using different SNR values, allowing the simulation to represent a wide range of wireless conditions.

SSDF attack model

The simulation includes both honest and malicious SUs. Malicious users manipulate their sensing reports using different attack strategies:

1. Energy inflation – artificially increases reported energy.
2. Energy deflation – artificially decreases reported energy.
3. Random perturbation – introduces random modifications to sensing values.

The research also introduces label noise, where some training labels are intentionally flipped. This represents imperfect or incorrect annotations that can occur in real-world datasets.

Proposed machine-learning methodology

The proposed detection process consists of five main stages:

1. Dataset generation – Generate sensing observations under different SNR and channel conditions.
2. Attack injection – Introduce malicious users and simulate SSDF attacks.
3. Feature engineering – Extract statistical characteristics such as mean, variance, and entropy from sensing behavior.
4. Model training – Train machine-learning classifiers, including Support Vector Machines and ensemble models, to distinguish honest and malicious users.
5. Evaluation – Assess the models using accuracy, precision, recall, and ROC-AUC on unseen data to determine their ability to generalize.

Key contribution

The main contribution is the development of a realistic and robust SSDF attack detection framework that considers SNR variability, probabilistic attacks, and labeling noise. Unlike many previous studies that use idealized datasets or deterministic attack behavior, this research attempts to reproduce the uncertainty of real wireless environments.

Conclusion

This paper compares machine learning models for detecting SSDF attacks in Cognitive Radio Networks. We created a simulation-based dataset to model spectrum sensing in different Signal-to-Noise Ratio scenarios and probabilistic adversarial behaviour. This work demonstrates the importance of cooperative spectrum sensing for improving reliability and spectrum utilization in Cognitive Radio Networks[10], [12]. We studied classification models with sensing data features, including Support Vector Machine, Random Forest, Logistic Regression and neural network-based approaches. Random Forest performed best among the tree-based methods in different scenarios. Experimental results confirm that intelligent learning-based detection methods can significantly enhance the security and robustness of cooperative spectrum sensing systems[8], [13]. The analysis shows that the models need to be able to capture non-linear relationships in the feature space to perform well. Linear decision boundaries, as provided by simpler models like Logistic Regression, are not sufficient to solve this problem. The high accuracy we obtained at the time suggests that the dataset remains separable between honest and malicious users. With changes in Signal-to-Noise Ratio and probabilistic attack behavior the feature distributions were distinct enough for models to classify with high precision.

References

[1] L. Duan, A. W. Min, J. Huang, and K. G. Shin, “Attack Prevention for Collaborative Spectrum Sensing in Cognitive Radio Networks,” Sep. 2011, doi: 10.1109/JSAC.2012.121009. [2] I. E. Ezhilarasi, J. C. Clement, and J. M. Arul, “A survey on cognitive radio network attack mitigation using machine learning and blockchain,” EURASIP J. Wirel. Commun. Netw., vol. 2023, no. 1, p. 98, Sep. 2023, doi: 10.1186/s13638-023-02290-z. [3] Yonghong Zeng and Ying-chang Liang, “Eigenvalue-based spectrum sensing algorithms for cognitive radio,” IEEE Transactions on Communications, vol. 57, no. 6, pp. 1784–1793, Jun. 2009, doi: 10.1109/TCOMM.2009.06.070402. [4] S. Haykin, “Cognitive radio: brain-empowered wireless communications,” IEEE Journal on Selected Areas in Communications, vol. 23, no. 2, pp. 201–220, Feb. 2005, doi: 10.1109/JSAC.2004.839380. [5] M. S. Hossain and M. S. Miah, “Machine learning-based malicious user detection for reliable cooperative radio spectrum sensing in Cognitive Radio-Internet of Things,” Machine Learning with Applications, vol. 5, p. 100052, Sep. 2021, doi: 10.1016/j.mlwa.2021.100052. [6] A. Taggu and N. Marchang, “Detecting Byzantine attacks in Cognitive Radio Networks?: A two-layered approach using Hidden Markov Model and machine learning,” Pervasive Mob. Comput., vol. 77, p. 101461, Oct. 2021, doi: 10.1016/j.pmcj.2021.101461. [7] G. V. P. Kumar and D. K. Reddy, “Hierarchical Cat and Mouse based ensemble extreme learning machine for spectrum sensing data falsification attack detection in cognitive radio network,” Microprocess. Microsyst., vol. 90, p. 104523, Apr. 2022, doi: 10.1016/j.micpro.2022.104523. [8] E. Ezhilarasi I and J. C. Clement, “Robust cooperative spectrum sensing in cognitive radio blockchain network using SHA-3 algorithm,” Blockchain: Research and Applications, vol. 5, no. 4, p. 100224, Dec. 2024, doi: 10.1016/j.bcra.2024.100224. [9] T. Yucek and H. Arslan, “A survey of spectrum sensing algorithms for cognitive radio applications,” IEEE Communications Surveys & Tutorials, vol. 11, no. 1, pp. 116–130, 2009, doi: 10.1109/SURV.2009.090109. [10] F. F. Digham, M.-S. Alouini, and M. K. Simon, “On the energy detection of unknown signals over fading channels,” in IEEE International Conference on Communications, 2003. ICC ’03., IEEE, pp. 3575–3579. doi: 10.1109/ICC.2003.1204119. [11] P. Kaligineedi, M. Khabbazian, and V. K. Bhargava, “Secure Cooperative Sensing Techniques for Cognitive Radio Systems,” in 2008 IEEE International Conference on Communications, IEEE, 2008, pp. 3406–3410. doi: 10.1109/ICC.2008.640.

Copyright

Copyright © 2026 Payal Dogra, Prof. Inderdeep Kaur Aulakh, Amanjot Kaur . This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.