<b>Federated Incremental Learning for IoT Device Identification</b>

<p dir="ltr">The rapid proliferation of Internet of Things (IoT) devices has intensified security vulnerabilities, necessitating robust methods for device identification. This thesis proposes a framework paradigm based on federated incremental learning, from which two algorithmic implementations are derived to address varying levels of IoT attack challenges. The proposed approach not only emphasizes IoT security through device identification, but also preserves client privacy and ensures scalability.</p><p dir="ltr">This framework is the first to combine federated learning and incremental learning to tackle practical IoT device identification problems. Through the federated learning mechanism, clients can share device knowledge collaboratively to enable cooperative defense, while avoiding privacy leakage caused by direct data sharing. By leveraging binary encoding and multi-channel stacking, FICDF effectively mitigates the catastrophic forgetting problem and achieves an average identification accuracy of approximately 94\% across multiple incremental tasks. Furthermore, it adopts k-means multi-centroid exemplar selection for data replay, which enables stable cross-task classification performance as new devices are continuously added.</p><p dir="ltr">In cross-network and adversarial environments, many packet-level features used for device fingerprinting become unreliable. An attention-based device identification model is proposed. A constrained set of communication features that remain stable under cross-network and attack scenarios is selected, enabling the generation of consistent manufacturer-level fingerprints across different networks and conditions.</p><p dir="ltr">Experiments conducted on real-world IoT communication datasets demonstrate that both frameworks achieve excellent device identification accuracy and strong cross-task scalability. The results validate that this federated incremental learning–based architecture offers a reliable and effective solution to enhance security in IoT networks.</p>