Graduation Year

2026

Document Type

Dissertation

Degree

Ph.D.

Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Electrical Engineering

Major Professor

Zhuo Lu, Ph.D.

Committee Member

Ismail Uysal, Ph.D.

Committee Member

Xiaomin Lin, Ph.D.

Committee Member

Attila Altay Yavuz, Ph.D.

Committee Member

Hao Zhou, Ph.D.

Keywords

federated learning, machine learning, machine unlearning, model poisoning, security

Abstract

With the rapid integration of Machine Learning (ML) into networking systems, ensuring the security and trustworthiness of these intelligent frameworks has become a paramount concern. While ML offers unprecedented capabilities in spectrum management and collaborative learning, it also introduces novel vulnerabilities that can be exploited by sophisticated adversaries. This dissertation investigates and addresses critical security challenges across three key dimensions of ML-driven networking: adversarial spectrum sensing, the security of privacy-preserving unlearning processes, and efficient post-attack model recovery.

First, we address the threat of adversarial spectrum attacks in cognitive radio networks, where malicious nodes manipulate sensing reports to disrupt spectrum access. We propose an efficient detection framework based on the Distance to the Decision Boundary (DDB). By developing a computationally optimized binary search method to estimate boundary distances, our approach achieves high detection accuracy with minimal overhead, providing a robust defense against input-layer perturbations in dynamic wireless environments.

Second, we explore the security risks inherent in Federated Unlearning (FU), a mechanism designed to comply with privacy regulations by removing specific data influence from global models. We identify a novel threat termed the FUsion backdoor, where adversaries exploit the unlearning process to inject malicious functionality. To counter this, we introduce SubID, a defense framework that utilizes subspace-based deviation identification to distinguish malicious model updates from benign noise, thereby securing the lifecycle of collaborative learning models.

Finally, we propose Federated Adversarial Unlearning (FAUN) to address the challenge of efficient model recovery after a poisoning attack is detected. Recognizing that traditional retraining is often computationally prohibitive in large-scale networks, FAUN leverages adversarial optimization to rapidly eliminate the influence of poisoned data. Our experimental results demonstrate that FAUN can restore model integrity in significantly fewer rounds than full retraining while maintaining high global accuracy.

In summary, this dissertation provides a comprehensive suite of methodologies and frameworks that enhance the robustness and resilience of machine learning within networking infrastructures. The proposed solutions contribute to the development of next-generation, trustworthy autonomous systems that are secure against both traditional and emerging adversarial threats.

Download

Share

COinS