Network Traffic Classification Under Label Scarcity and Privacy Constraints Using Federated Self-Supervised and Confident Learning
Abstract
Network Traffic Classification (NTC) is essential for network management and security; however, it faces challenges such as label scarcity, pseudo-label noise, privacy concerns, non-independent and non-identically distributed (non-IID) data, and class imbalance in distributed environments. This thesis proposes two interconnected frameworks to address these issues. The centralized approach integrates traffic-adapted self-supervised learning (SSL) with confident learning (CL), employing a constraint-consistent autoencoder (AE) and enhanced tabular contrastive learning (TabCL) for robust pseudo-label generation from unlabeled flows, followed by CL-based denoising using per-class quantile thresholds, calibration-aware weighting, and balanced retention. This centralized SSL+CL pipeline is proven to be highly effective, establishing a strong performance benchmark. The federated extension, FedSSL-NTC, distributes this process across clients, incorporating FedProx for non-IID stability, class-weighted losses for imbalance, sample-size-weighted (SSW) Federated Averaging (FedAvg) for fair aggregation, and a tailored secure aggregation (SecAgg) protocol for privacy preservation, ensuring no raw data sharing. Evaluations on a self-generated, unlabeled dataset combined with ISCX VPN-nonVPN and the UCDavis–QUIC benchmark demonstrate high effectiveness. Experimental results demonstrate that FedSSL-NTC achieves near-centralized accuracy. These contributions advance label-efficient,…
Read the paper
Related studies
- Active Learning during Federated Learning for Object Detection
- Fairness-aware federated learning
- Decentralized Federated Learning
- Building a Personally Identifiable Information Recognizer in a Privacy Preserved Manner Using Automated Annotation and Federated Learning
- Preservation of Patient Level Privacy: Federated Classification and Calibration Models
- Efficient explainable federated learning
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.