Federated Learning (FL) has emerged as a decentralized technique, where contrary to traditional centralized approaches, devices perform a model training in a collaborative manner, while preserving data privacy. Despite the existing efforts made in FL, its environmental impact is still under investigation, since several critical challenges regarding its applicability to wireless networks have been identified. Towards mitigating the carbon footprint of FL, the current work proposes a Genetic Algorithm (GA) approach, targeting the minimization of both the overall energy consumption of an FL process and any unnecessary resource utilization, by orchestrating the computational and communication resources of the involved devices, while guaranteeing a certain FL model performance target. A penalty function is introduced in the offline phase of the GA that penalizes the strategies that violate the constraints of the environment, ensuring a safe GA process. Evaluation results show the effectiveness of the proposed scheme compared to two state-of-the-art baseline solutions, achieving a decrease of up to 83% in the total energy consumption.

Network traffic anomalies can have a detrimental effect on end-to-end network performance and reliability, compromising severely (demanding) services and applications, such as those offered by the fifth-generation (5G) mobile networks and beyond. To prevent the network drift towards increasingly inefficient operating modes, such traffic anomalies need to be detected first, so that the proper actions be taken to avoid such undesirable network drift. This paper proposes a novel framework focusing on proactively detecting such traffic anomalies, that is, predicting upcoming network traffic anomalies before they actually occur. To this end, two mechanisms are developed: a mechanism aiming at identifying different network traffic behaviors in an automated manner; and a mechanism for predicting network traffic behaviors for the next time interval spanning several seconds. The automated network traffic anomaly identification is realized via the application of clustering and decision tree-based learning. A time series model (specifically, a Bidirectional Long Short Term Memory (BiLSTM) Autoencoder) is employed for the proactive detection of forthcoming network traffic behaviors. Evaluation results are derived showing a prediction accuracy of up to 90.02%, demonstrating the effectiveness and viability of the proposed framework, as well as its potential for higher effectiveness compared to a state-of-the-art solution.

Beyond 5G networks bring a new era in system automation, by introducing new and demanding, in terms of Quality of Service (QoS), use cases and applications. Predicting the QoS for end users in a timely manner and enabling service adaptation methods to react in advance in case of QoS degradation is of high importance, especially for safety-critical applications such as in vehicular communications. Current state-of-the-art approaches propose solutions towards the identification of potential QoS deterioration in a centralized manner. However, centralized solutions may raise privacy issues, since sensitive user information may need to be transmitted to communication network entities for processing and analysis. Other practical limitations of centralized solutions may also arise, such as the computational bottleneck and the fast increase of signaling overhead with number of end users. This study proposes a distributed QoS prediction scheme based on the well-known Long Short-Term Memory (LSTM) architecture to account for the natural high correlation of samples closely located in time. The primary target of the proposed scheme is to provide accurate QoS predictions up to several seconds, while preserving data privacy and reducing signaling overheads related to the exchange of information between the involved nodes. The evaluation of the proposed scheme indicates its potential gains and effectiveness compared to centralized state-of-the-art OoS prediction solutions.

In the era of Industrial Internet of Things (IIoT) and Industry 4.0, an immense volume of heterogeneous network devices will coexist and contend for shared network resources, in order to satisfy the very challenging IIoT applications, requiring ultra-reliable and ultra-low latency communications. Although novel key enablers, such as Network Slicing, Software Defined Networking (SDN) and Network Function Virtualization (NFV) have already offered significant advantages towards more efficient and flexible network and resource management approaches, the particular characteristics of IIoT applications pose additional burdens, mainly due to the complex wireless environments, high number of heterogeneous network devices, sensors, user equipments (UEs), etc., which may stochastically demand and contend for the - often scarce - computing and communication resources of industrial environments. To this end, this paper introduces PRIMATE, a novel, Artificial Intelligence (AI)-driven framework for the profiling of the networking behavior of such UEs, devices, users and things, which is able to operate in conjunction with already standardized or forthcoming, AI-based network resource management processes towards further gains. The novelty and potential of the proposed work lies on the fact that instead of attempting to either predict raw network metrics in a reactive manner, or predict the behavior of specific network entities/devices in an isolated manner, a big data-driven classification approach is introduced, which models the behavior of any network device/user from both a macroscopic, as well as service-specific perspective. The extended evaluation at the last part of this work shows the validity and viability of the proposed framework.