Traffic classification plays a crucial role in optimizing network management, enhancing security, and enabling intelligent resource allocation in distributed network systems. However, traditional Federated Learning (FL) approaches struggle with domain heterogeneity, as network traffic characteristics vary significantly across different domains due to diverse infrastructure, applications, and usage patterns. This results in degraded performance when applying a single global model across all domains. To overcome this challenge, we propose KD-PFL-TC, a Knowledge Distillation-based Personalized Federated Learning framework for Traffic Classification, aimed to balance global knowledge sharing with personalized model adaptation in heterogeneous network environments. Our approach leverages knowledge distillation to enable collaborative learning without directly sharing raw data, preserving privacy while mitigating the negative effects of domain shifts. Each domain refines its local model by integrating insights from a global model and peer domains while maintaining its unique traffic distribution. To further enhance performance, we introduce an adaptive distillation strategy that dynamically adjusts the influence of global, peer, and local knowledge based on the similarity between traffic distributions, ensuring optimal knowledge transfer designed to each domain’s characteristics. Extensive experiments on real-world traffic datasets show that KDPFLTC maintains 88.0% accuracy under high heterogeneity (vs. 75.0% for FedAvg) while reducing communication overhead by 60%, delivering an efficient and robust solution for large-scale, heterogeneous networks.

Encrypted network traffic classification has become a critical task with the widespread adoption of protocols such as HTTPS and QUIC. Deep learning-based methods have proven to be effective in identifying traffic patterns, even within encrypted data streams. However, these methods face significant challenges when confronted with new applications that were not part of the original training set. To address this issue, knowledge transfer from existing models is often employed to accommodate novel applications. As the complexity of network traffic increases, particularly at higher protocol layers, the transferability of learned features diminishes due to domain discrepancies. Recent studies have explored Deep Adaptation Networks (DAN) as a solution, which extends deep convolutional neural networks to better adapt to target domains by mitigating these discrepancies. Despite its potential, the computational complexity of discrepancy metrics, such as Maximum Mean Discrepancy, limits DAN’s scalability, especially when applied to large datasets. In this paper, we propose a novel DAN architecture that incorporates Smooth Characteristic Functions (SCFs), specifically SCF-unNorm (Unnormalized SCF) and SCF-pInverse (Pseudo-inverse SCF). These functions are designed to enhance feature transferability in task-specific layers, effectively addressing the limitations posed by domain discrepancies and computational complexity. The proposed mechanism provides a means to efficiently handle situations with limited labeled data or entirely unlabeled data for new applications. The aim is to limit the target error by incorporating a domain discrepancy between the source and target distributions along with the source error. Two statistics classes, SCF-unNorm and SCF-pInverse, are used to minimize this domain discrepancy in traffic classification. The experimental results demonstrate that our proposed mechanism outperforms existing benchmarks in terms of accuracy, enabling real-time traffic classification in network systems. Specifically, we achieve up to 99% accuracy with an execution time of only three milliseconds in the considered scenarios.

Server and route selection (SARS) optimization is a critical aspect of traffic engineering to allocate network resources to meet diverse service requirements effectively. Existing studies have primarily focused on finding profitable or optimal solutions for the SARS problem within current time steps, considering specific constraints. However, they often have failed to address the dynamic and uncertainty of future network states. To address this gap, this paper proposes an algorithm named GAL to optimize server costs and response time while accounting for future network dynamics. GAL combines a server selection inspired by the gambling theory and a network routing based on Long Short-Term Memory Networks (LSTM). The server selection method is formulated as a gambling problem and solved using the decision-making Tug-of-War (TOW) dynamic algorithm. The routing mechanism is optimized based on predictions of future network states made by LSTM neural networks, which excel in capturing long-term dependencies. We have implemented GAL through a distributed software-defined networking (SDN) system and obtained good evaluation results regarding average response time and server cost compared to benchmark methods. These results demonstrate that GAL can effectively tackle the SARS optimization problem by considering present constraints and future network dynamics. This study can advance traffic engineering and lays a foundation for more robust resource allocation strategies in dynamic network environments.

Software-defined networking (SDN) is a novel net-working paradigm for network management and autonomous systems. However, SDN has some challenges with scalability and quality of services (QoS) in distributed multi-domain scenarios due to the unprecedented growth of heterogeneous characteristics services. There is a current gap in a standard routing mechanism for satisfying various service requirements in distributed SDN. Most existing works design a homogeneous routing strategy for heterogeneous services, which might need to be more scalable and efficient for the future of rising heterogeneous online services. This study proposes a multi service-oriented routing mechanism for multi-domain SDN, which aims to help Internet service providers (ISPs) achieve high QoS and service-level agreements (SLAs). The mechanism utilizes a service classification (through a deep learning model) and optimizes network routing (using a new cost function containing both QoS and the server load). The mechanism has been integrated into the Knowledge-defined heterogeneous network architecture and tested on four prevalent considered services: E-commerce, Interactive Data, Video On-demand, and Bulk Data Transfer. The experimental results indicate that the proposed service-oriented routing mechanism outperforms the benchmark in terms of faster server response time while reducing up to 25% of the network congestion.

Network traffic management is crucial in Computer Networks and the Internet of Things. Indeed, classifying network traffic is the foundation for enhancing the quality of management mechanisms. However, traditional traffic classification methods, such as port-based, deep packet inspection, and statistic-based, are limited in identifying new encrypted traffic characteristics. Deep Learning-based classification approaches that consider packet-based features have been explored to address this challenge. Along with other deep learning methods, Transfer Learning, where a new model can inherit knowledge previously learned by a base model, is commonly used to increase classification performance in low data resources. Unfortunately, feature transferability may decline in transfer learning. This paper proposes an encrypted traffic classification mechanism that leverages the Deep Adaptation Network architecture with Mean Embedding Test to overcome this limitation. Our experimental results show that the proposed mechanism surpasses existing benchmarks’ accuracy and can classify encrypted traffic in real-time.