Abstract
Federated Learning (FL) has garnered significant interest recently due to its potential as an effective solution for tackling many challenges in diverse application scenarios, for example, data privacy in network edge traffic classification. Despite its recognized advantages, FL encounters obstacles linked to statistical data heterogeneity and labeled data scarcity during the training of single-task models for machine learning-based traffic classification, leading to hindered learning performance. In response to these challenges, adopting a hard-parameter sharing multi-task learning model with auxiliary tasks proves to be a suitable approach. Such a model has the capability to reduce communication and computation costs, navigate statistical complexities inherent in FL contexts, and overcome labeled data scarcity by leveraging knowledge derived from interconnected auxiliary tasks. This paper introduces a new framework for federated auxiliary hard-parameter sharing multi-task learning, namely, FedAuxHMTL. The introduced framework incorporates model parameter exchanges between edge server and base stations, enabling base stations from distributed areas to participate in the FedAuxHMTL process and enhance the learning performance of the main task-network edge traffic classification. Empirical experiments are conducted to validate and demonstrate the FedAuxHMTL's effectiveness in terms of accuracy, total global loss, communication costs, computing time, and energy consumption compared to its counterparts.
Introduction
Classifying network traffic at the edge—identifying whether a stream is Google Music, YouTube, or a file download—is essential for maintaining Quality of Service (QoS) and security. However, doing this effectively in a decentralized environment faces three major roadblocks: data privacy, statistical heterogeneity (Non-IID data), and labeled data scarcity. While Federated Learning (FL) addresses the privacy issue by keeping data local, it often struggles with the latter two points.
FedAuxHMTL is a novel framework that combines FL with Multi-Task Learning (MTL) to overcome these hurdles. By sharing "hard parameters" across a main task and several auxiliary tasks, the system achieves higher accuracy with significantly less communication overhead than standard methods.
Methodology: FedAuxHMTL Framework
The core of FedAuxHMTL is its architecture and the strategic balancing of different learning objectives.
1D-CNN MTL Architecture
The local model is a 1D Convolutional Neural Network (CNN) utilizing Hard-Parameter Sharing. In this setup, the initial layers (feature extractors) are shared across all tasks, while the final "heads" are task-specific. This architecture forces the model to learn robust, generalized features of network traffic that are useful across multiple domains.
Auxiliary Tasks and Loss Weighting
To bolster the main classification task, the model simultaneously learns to classify Flow Duration and Bandwidth. These auxiliary tasks provide extra "hints" to the model, acting as a regularizer that helps it learn even when labels for the main task are sparse.
Balancing these tasks is achieved through Random Loss Weighting (RLW). Unlike Equal Loss Weighting (ELW), RLW assigns dynamic weights to auxiliary tasks from a distribution, normalized via a Softmax function. This prevents any single task from dominating the learning process and helps the model explore the loss landscape more effectively.
Experimental Setup
The framework was evaluated using the QUIC dataset, focusing on encrypted traffic from five Google services: Google Drive, Search, Music, Docs, and YouTube. The researchers partitioned 90% of the 6,439 samples for training and 10% for testing. Crucially, the data was distributed across six base stations in a non-IID manner to simulate a realistic, fragmented edge network.
Key Results and Discussion
Empirical results demonstrate that FedAuxHMTL significantly outperforms standard Federated Averaging (FedAvg) and other multi-task baselines.
- Faster Convergence: FedAuxHMTL with RLW reached 80% accuracy on the main task in just 29 communication rounds, compared to 62 rounds for standard FedAvg.
- Efficiency Gains: The framework achieved a 53% reduction in communication costs and a 32% reduction in computing time.
- Energy Savings: Due to faster convergence, energy consumption was reduced by over 50%—roughly 76 Joules for FedAuxHMTL versus 163 Joules for FedAvg to reach the same accuracy level.
The most striking insight is that auxiliary tasks act as "data boosters." Learning to predict flow duration helps the model understand the underlying structure of traffic, which indirectly boosts main task classification accuracy without requiring more main-task labels.
Conclusion and Future Work
FedAuxHMTL provides a robust solution for network edge traffic classification by leveraging the synergy between Federated Learning and Multi-Task Learning. While the current model assumes homogeneous hardware across base stations, future work will address system heterogeneity and explore more diverse, large-scale datasets. The authors also plan to provide a deeper theoretical analysis of the mathematical bounds of Federated Multi-Task Learning performance gains.
Summary updated