Federated Learning Improves Distributed Spectrum Sensing for Future Wireless Networks

The increasing demand for wireless connectivity, fuelled by advancements in 5G and the development of 6G, is intensifying the problem of spectrum scarcity, a critical limitation for mobile networks. Ngoc Duy Pham, Thusitha Dayaratne, and Carsten Rudolph, from Monash University, along with colleagues from Swinburne University of Technology, Sharif University, CSIRO’s Data61, and The University of Melbourne, present a new approach to dynamic spectrum allocation that overcomes the limitations of traditional methods. Their research focuses on federated learning, a distributed machine learning technique, to enable spectrum sensing without requiring centralised data collection and addressing key vulnerabilities to malicious attacks. The team demonstrates a semi-supervised learning method that achieves high accuracy even with limited labelled data, and introduces a novel defence mechanism, inspired by vaccination, that effectively protects against data poisoning attacks, ensuring robust and reliable spectrum sensing in challenging real-world conditions.

Federated Learning Solves Spectrum Scarcity Challenges

The relentless growth of wireless communication, driven by advancements towards 5G and the emerging 6G technologies, is creating a critical challenge: spectrum scarcity. As more devices connect and demand bandwidth, efficiently sharing available radio frequencies becomes paramount. Dynamic spectrum allocation, the ability to sense and dynamically share unused frequencies, offers a promising solution, but traditional methods struggle with accuracy and reliability. Machine learning offers a potential leap forward, enabling more precise spectrum sensing, but centralized approaches require transmitting sensitive user data, raising privacy concerns and creating significant communication burdens.

Federated learning has emerged as a compelling alternative, allowing devices to collaboratively train machine learning models without directly sharing raw data. However, practical implementation faces hurdles, notably the scarcity of labeled data needed to train effective models. Researchers have developed an approach that combines semi-supervised federated learning with energy detection, enabling accurate spectrum sensing even with limited labeled data and reducing the need for extensive manual effort. Beyond data limitations, federated learning systems are vulnerable to malicious attacks, where compromised devices can poison the training process.

Existing defenses often rely on the assumption that the majority of participating devices are trustworthy, a flawed premise in open networks. To address this, scientists have developed a novel defense mechanism inspired by vaccination, strengthening the system against attacks without relying on majority-based assumptions and ensuring robust performance even with a significant proportion of compromised devices. This research represents a significant step towards realizing the full potential of dynamic spectrum allocation in future wireless networks, paving the way for more efficient, private, and resilient spectrum sensing.

Federated Learning with Semi-Supervised Energy Detection

Researchers developed a novel approach to spectrum sensing that addresses the challenges of increasing wireless device density and limited available spectrum. Recognizing the limitations of traditional methods, the team focused on federated learning, a technique allowing devices to collaboratively train a model without sharing raw data. A key innovation lies in overcoming the difficulty of obtaining sufficient labeled data for training these models. They introduced SEMISS, which combines federated learning with semi-supervised learning and energy detection, allowing the system to learn from both labeled and unlabeled data and significantly improving efficiency.

SEMISS utilizes energy detection to enable devices to self-correct labels on unlabeled data, effectively creating a locally refined dataset and achieving sensing accuracy comparable to that of fully labeled datasets. Beyond data limitations, the researchers also addressed security vulnerabilities inherent in distributed learning systems, specifically the threat of data poisoning attacks. Their analysis revealed that existing defenses are easily circumvented when malicious devices collaborate. To counter this, they developed SSVAX, a defense mechanism inspired by vaccination. Unlike traditional approaches, SSVAX proactively strengthens the system by simulating attacks and generating ‘vaccines’ that identify and neutralize malicious model updates. Through both synthetic and real-world testing, the team demonstrated that their combined approach achieves near-perfect accuracy on unlabeled data and maintains resilience against sophisticated data poisoning attacks, paving the way for more secure and efficient wireless networks.

Federated Learning Enables Private Spectrum Sharing

Researchers are tackling the growing challenge of spectrum scarcity in wireless communication with a novel approach to dynamic spectrum allocation, leveraging distributed machine learning. As the number of wireless devices continues to increase with the advent of 5G and the development of 6G technologies, efficient spectrum management is becoming ever more critical. This work introduces a system where devices collaboratively learn to share spectrum without directly exchanging sensitive data, addressing privacy concerns and bandwidth limitations. The team focused on federated learning, where individual devices train local models and share only model improvements with a central server.

A key innovation lies in overcoming the difficulty of obtaining sufficient labeled data for training these models. By combining federated learning with semi-supervised learning, the system can effectively learn from both labeled and unlabeled data, significantly improving performance in practical scenarios where labeled data is scarce. Furthermore, the researchers addressed a critical security vulnerability in these distributed learning systems: data poisoning attacks. Malicious devices could potentially compromise the entire system by submitting deliberately flawed model updates. Existing defenses proved inadequate against sophisticated attacks.

To counter this, the team developed a novel defense mechanism inspired by vaccination, where the system is proactively inoculated against malicious data. Extensive testing on both simulated and real-world datasets demonstrates the robustness and effectiveness of this new system. Results show that the federated learning approach, combined with semi-supervised learning and the vaccination-inspired defense, achieves high accuracy and maintains resilience against attacks, even with a substantial number of malicious participants. This represents a significant step forward in enabling efficient, secure, and privacy-preserving spectrum allocation for the next generation of wireless networks.

Vaccination Defends Federated Spectrum Sensing Systems

This work addresses critical challenges in dynamic spectrum allocation, a key solution to increasing demand for wireless spectrum in the 5G and future 6G eras. Researchers developed methods to improve federated learning-based spectrum sensing, which allows devices to collaboratively learn spectrum availability without sharing sensitive data. The team tackled the problem of limited labeled data by combining semi-supervised learning with energy detection, enabling accurate spectrum sensing even with predominantly unlabeled datasets. Furthermore, the research investigated vulnerabilities to data poisoning attacks in these systems and proposed a novel defense mechanism inspired by vaccination.

This approach effectively mitigates malicious interference without relying on traditional majority-based defenses, demonstrating robustness against both targeted and untargeted attacks, even with a significant number of compromised participants. Future work could focus on developing more versatile defenses, and the potential impact of jammers warrants further investigation. The findings validate the effectiveness and security of these methods for active spectrum sensing in current and future radio networks, paving the way for more efficient and reliable wireless communication.