عمار المترفي

The move to sixth generation (6G) networks brings about a new level of attack surface that is dynamic an‎d characterized by signaling storms an‎d adversarial intrusions, making it impossible for the traditional centralized Intrusion Detection System (IDS) to be effective, especially due to ban‎dwidth limitations an‎d the need to comply with privacy regulations such as GDPR. This thesis proposes an‎d analyzes a Federated Autoencoder-based Anomaly Detection Framework that moves security intelligence to the edge of the network, allowing for joint model training without sharing raw user data. The main contribution of this research work is the combination of an unsupervised deep Autoencoder model with a new Adaptive Aggregation method tailored to address the Non-Independent an‎d Identically Distributed (Non-IID) data that is typical of mobile networks. Experimental verification on the ToN-IoT an‎d 5G-NIDD datasets proved the effectiveness of the framework in detecting complex threats in a heterogeneous simulation environment. The model had a near-perfect Recall of 1.00, correctly detecting 99.996% of attack occurrences, with an Accuracy of 93.68% an‎d an F1-Score of 0.9602. Additionally, the global model proved to be highly robust, achieving a training loss of 2.501 with large statistical fluctuations in the federated clients. These findings clearly support the hypothesis that decentralized, unsupervised learning can provide detection capabilities on par with centralized approaches, while also providing the robust privacy guarantees that are essential for secure 6G communication.

شبكه‌هاي 6G , يادگيري فدرال

6G Networks , Federated learning

Ammar Mutrafi

Dr. Diyanat