Silent-App-Aware Federated Machine Unlearning for Encrypted Network Traffic Classification

In recent years, federated learning-based network traffic classification has emerged as a promising Artificial Intelligence (AI) approach for encrypted traffic identification, enabling accurate classification while preserving data privacy. To meet the rising demand for the right to be forgotten, Federated Machine Unlearning (FMU) has been introduced as an AI-driven forget ting paradigm. However, conventional FMU methods overlook silent applications removed due to guideline violations, resulting in reduced classifier stability and limited aggregation scalability. To address these challenges, this paper proposes FedUnSilApp, short for FMU for silent applications, a novel FMU framework that integrates a learning module, a forgetting module, and a scalable aggregation strategy. In particular, the forgetting module prunes neurons associated with silent applications in the classifier, minimizing their residual influence. The scalable aggregation strategy separately processes discriminative layers of clients requesting forgetting and those of other clients, thereby enhancing both stability and scalability of the global model. In addition, FedUnSilApp is applied to the real-world scenario of encrypted network traffic classification, enabling compliance with deletion requests while maintaining robust classification performance. The effectiveness of FedUnSilApp is validated through experiments, which show that FedUnSilApp outperforms state-of-the-art methods, improving accuracy by over 6% on the NJUPT dataset.