Federated Learning for Network Traffic Classification: A Knowledge Consolidation Approach
Hai Anh Tran, Huỳnh Thị Thanh Bình, Abdelhamid Mellouk · IEEE Transactions on Network Science and Engineering · 2025
Thegrowing complexity of distributed systems and dynamic environments has presented significant challenges for incremental learning, particularly in handling non-IID data distributions and addressing resource constraints. Traditional federated learning approaches often suffer from limitations such as catastrophic forgetting, high communication overhead, and misalignment between global and local knowledge, reducing their effectiveness in such scenarios. To overcome these challenges, we propose the Incremental Knowledge Consolidation (IKC) framework, which introduces a novel combination of BERT-based feature embeddings and consolidated attention maps to enable adaptive, class-incremental learning in distributed network domains. The key contributions of our framework are threefold. First, IKC extracts contextual, high-dimensional embeddings from local traffic instances using BERT, capturing nuanced network behaviors while ensuring scalability across distributed devices. Second, it introduces a dynamic knowledge distillation mechanism that shares distilled prototypes and attention maps across devices, facilitating global-local alignment without centralizing raw data. Third, IKC incorporates an adaptive classification mechanism, leveraging global prototypes and attention maps to prioritize relevant traffic features and dynamically adapt to newly emerging traffic types. Experimental results demonstrate IKC's superiority over benchmark methods, including FedAvg, FedProx, GEM, and SCAFFOLD. IKC outperforms these benchmarks in classification accuracy, particularly for minority traffic classes, and achieves a compression ratio of over 90%, significantly reducing communication overhead. Additionally, IKC maintains high global-local alignment, as evidenced by its superior cosine similarity scores, ensuring consistent learning across devices in non-IID environments. These findings highlight IKC's potential as an efficient and robust framework for incremental learning in distributed systems.