Back to search

École de technologie supérieure

Federated learning empowered reliable jamming detection in the context of heterogeneous wireless ad-hoc networks

Abstract

dc:description

Today, data represent a key element for training and improving the performance of intelligent systems. However, sharing such data over the network can be risky because of their sensitivity and confidentiality. In this context, tactical wireless networks (TWNs) are widely used in military operations to share crucial information among nodes deployed on the battlefield. These networks use various radio waveforms, such as those employed in Wireless Local Area Network (WLAN), Long-Term Evolution (LTE), and Fifth Generation (5G) networks, which introduces significant data heterogeneity across the network. In addition, TWNs operate in hostile environments where normal network operations can be vulnerable to various types of jamming. Therefore, jamming detection is a crucial step to prevent attacks in tactical networks. Traditionally, methods such as cooperative detection and energy detectors have been used to identify anomalies. However, these conventional methods have significant limitations in terms of accuracy and robustness in increasingly complex environments. In response to these challenges, waveform recognition has been widely used to identify interference. Yet, while several deep learning methods have been developed for waveform recognition, the need to preserve data confidentiality limits the sharing of this information, thereby degrading the performance of centralized approaches. Moreover, centralized methods suffer from scalability issues and single points of failure, making networks more vulnerable to targeted attacks. In this context, the present thesis develops a jamming detection and waveform classification system for tactical networks while ensuring a trade-off between learning performance and data confidentiality. To this end, we consider a large-scale heterogeneous distributed network supporting different types of signals. We propose federated learning (FL) based on the spectral correlation function (SCF), which is used as a feature extraction method for jamming detection and waveform classification. The proposed FL frameworks keep raw training data local, thereby reducing direct data exposure while avoiding the need to centralize sensitive information. In addition, we analyze the performance of our solutions in the presence of intelligent jammers, enabling network nodes to establish secure communication links and avoid jammers. These solutions make TWNs more robust and reliable, thereby increasing their operational efficiency in hostile environments. Chapter 3 presents the modeling and implementation of a distributed TWN using heterogeneous waveforms. Specifically, we first generate a dataset based on practical parameters, including channel fading, path loss, and interference. Next, an FL-based solution is implemented for jamming detection and waveform classification in distributed TWNs. This solution uses SCF estimation to extract signal features and employ them during the training and testing processes. In addition, a comparison of the extracted features with those from in-phase/quadrature (I/Q) components is conducted. The proposed solution is then compared with existing approaches. Finally, several experiments are conducted to evaluate the performance of the proposed solution in terms of accuracy, complexity, execution time, and data privacy protection. In Chapter 4, we introduce a novel FL algorithm, named AAT-Fed, specifically designed for stochastic, distributed terrestrial and non-terrestrial tactical networks (SDT-TNT). Unlike the solution previously proposed in Chapter 3, which used SCF for signal feature extraction, the approach developed in this chapter goes further by integrating convolutional variational autoencoders (C-VAEs) to detect jammers without prior knowledge. The AAT-Fed algorithm addresses challenges posed by data heterogeneity and the non independent and identically distributed (non-IID) nature of signals in these networks. It provides improved robustness and detection accuracy under severe jamming conditions, particularly at low jammer-to-noise ratios (JNRs). An in-depth performance analysis shows that AAT-Fed outperforms conventional FL approaches, such as FedAvg and FedProx, as well as centralized methods, thereby ensuring greater efficiency in complex and dynamic scenarios. In Chapter 5, we introduce an innovative approach for 6G networks that integrates terrestrial infrastructure for their reliability and speed with non-terrestrial networks for their wide coverage, thereby creating a robust and seamless communication network. This chapter also addresses adaptive jamming in uplink (UL) and downlink (DL) communications. A 3D simulation framework is developed to model these communications in stochastic integrated terrestrial and non-terrestrial networks (SI-TNTNs), using a stochastic node distribution based on the homogeneous Poisson point process (HPPP). We also propose the FedPMA algorithm, which combines local encoders for feature extraction using SCF with a probabilistic aggregator based on a multivariate normal (MVN) distribution to improve performance and data confidentiality. Experimental results show that this approach significantly improves model convergence, reduces communication costs, and effectively prioritizes client updates, even in environments subject to adaptive jamming. In Chapter 6, we address the complex challenges of SAG-ISAC networks, where the coexistence of sensing and communication functionalities is compromised by dual-threat jammers. Unlike previous approaches, which mainly focused on static or adaptive jammers, this chapter considers sophisticated jammers capable of simultaneously disrupting user equipment (UE) communications and compromising UAV sensing functions by manipulating sensing signals. To address these challenges, we introduce the ReFDRL framework, which integrates deep reinforcement learning (DRL) and FL to dynamically adjust the weights assigned to local model updates. This framework uses vision transformers (ViTs), in which UAVs exploit features obtained from time-frequency distributions (TFD) for sensing, while UEs use SCF-based features for communications. A key innovation lies in the use of a deep Q-learning (DQL)-based mechanism to dynamically adjust client contributions according to the performance of their local models. Experimental results show that ReFDRL effectively (i) detects dual-threat jamming with greater accuracy than existing methods; (ii) optimizes the trade-off between sensing and communication tasks; and (iii) efficiently handles heterogeneous and non-IID data, thereby improving the performance of SAG-ISAC networks.

Degree

thesis:*
Name thesis:degree_name
Thèse de doctorat électronique
Level thesis:degree_level
phd
Grantor dc:publisher
École de technologie supérieure
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Meftah, Aida

Subjects

dc:subject × 1

Rights

Language dc:language
en

Identifiers

dc:identifier.*
Identifier
INSERT-YOUR-LIBRARY-ID-HERE4145

Chain of custody

source
Harvested from
ETS (Quebec)
Base URL
espace.etsmtl.ca/cgi/oai2
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Meftah, Aida. Federated learning empowered reliable jamming detection in the context of heterogeneous wireless ad-hoc networks. phd thesis, École de technologie supérieure, 2026.