Optimisation des réseaux IoT via apprentissage fédéré sécuritaire en edge computing

Thesis Abstract

The rapid proliferation of Internet of Things (IoT) deployments in industrial and urban environments has intensified concerns over data privacy, communication efficiency, and resilience against adversarial threats, motivating the exploration of secure federated learning (FL) in edge computing as a governance mechanism for distributed model updates. This study addresses the challenge of optimizing IoT network performance while preserving data sovereignty and mitigating model-poisoning risks in heterogeneous edge environments. The aim is to develop a scalable FL-enabled edge architecture that enhances network throughput, reduces latency, and sustains high classification accuracy for anomaly detection and traffic routing under stringent privacy and security constraints. Specific objectives include (i) modeling an edge-centric FL framework with secure aggregation and differential privacy, (ii) evaluating communication overhead and convergence behavior across diverse IoT topologies, (iii) benchmarking anomaly-detection accuracy and energy efficiency against centralized and non-federated baselines, (iv) analyzing resilience to Byzantine and data-poisoning attacks, and (v) providing guidelines for deployment in real-world smart-city and industrial IoT settings. The research adopts a mixed-methods design, combining quantitative simulation-based experiments with qualitative expert feedback. The population comprises IoT devices and edge servers distributed across three representative topologies a smart factory, a campus environment, and a metropolitan sensor network. A stratified sample of 120 edge nodes and 30 edge servers is selected, with 60 randomly assigned to the secure FL condition and 60 to traditional federated learning without secure aggregation. Data collections use synthetic yet realistic IoT traces generated via the iFogSim toolkit, complemented by publicly available datasets for intrusion detection (CICIDS2017) and device-level energy consumption. Instrumentation includes (i) a secure FL protocol implemented atop a lightweight secure aggregation scheme, (ii) a traffic-annotation module for labeling anomalous patterns, and (iii) a measurement suite for latency, bandwidth usage, energy consumption, and model accuracy. Validity and reliability are established through instrument calibration, pilot testing on a closed testbed, and repeated runs (n=30 per topology) to ensure statistical robustness. Data analysis proceeds in three phases. First, descriptive statistics and inferential analyses assess network performance indicators (throughput, end-to-end latency, packet loss) and model performance (precision, recall, F1-score) under varying data heterogeneity and participation rates. Second, regression analyses (hierarchical and ridge) examine the relationship between federation parameters (aggregation frequency, differential privacy budgets, and secure-m aggregation overhead) and performance metrics. Third, resilience is evaluated via robustness tests against simulated Byzantine and data-poisoning attacks, employing ANOVA and non-parametric equivalents to compare secure FL against baselines. Additionally, thematic analysis of expert feedback from system engineers and security analysts informs practical constraints and deployment prioritizations. The study integrates theoretical foundations from Federated Learning theory, Secure-Multiparty Computation, and Edge Intelligence, referencing the Bayesian learning paradigm to model uncertainty, and the Information Bottleneck principle to balance privacy and utility. Expected findings indicate that the secure FL framework will achieve superior privacy preservation (local differential privacy with ? in the range [1, 3]), while maintaining comparable global model accuracy to centralized training within a 2–4% margin, particularly in data-scarce or highly skewed partitions. It is anticipated that secure aggregation will introduce modest communication overhead, increasing per-round latency by approximately 8–15% but yielding substantial gains in resilience to poisoning attacks (mean F1 improvement of 6–9% under attack scenarios). The analysis is expected to reveal topology-dependent trade-offs factory-like environments with stable connectivity favor higher aggregation frequency, whereas city-scale networks benefit from adaptive participation and dynamic QoS-aware routing to sustain performance. The study contributes to knowledge by providing a rigorously evaluated design for secure FL in edge-empowered IoT ecosystems, outlining a pragmatic blueprint for deployment that integrates differential privacy, secure aggregation, and attack-resilient learning. It advances theoretical understanding of privacy-utility-security trade-offs in distributed edge learning and informs policy and industry standards for secure IoT analytics. Practical recommendations include guidelines for selecting privacy budgets, aggregation schedules, and resource provisioning to optimize performance while maintaining robust security guarantees. The main conclusion posits that secure federated learning at the edge is a viable solution for IoT network optimization, delivering measurable improvements in efficiency and resilience without compromising data sovereignty, with specific attention to topology-aware configuration and adaptive defense mechanisms.