Fine-tuning large pre-trained speech recognition models on resource-constrained edge devices within federated learning frameworks presents significant challenges due to computational limitations, communication overhead, and the complexities of mixed-precision training. To address these issues, we propose FedMP-Head, a two-step method designed to optimize training efficiency while maintaining high accuracy. In the first step, we freeze the encoder of the foundation model and train only a lightweight classifier head before applying mixed-precision quantization. This approach reduces the size of global model updates to approximately 780K parameters and maintains the accuracy during the downstream task of classification. In the second step, we implement budgeted layer-wise gradient scaling and correction with mixed-precision training on the client side. This technique optimizes computational cost and memory footprint on resource-constrained clients by effectively utilizing gradients, even with limited precision and computational capabilities. FedMP-Head accelerates training, reducing convergence time compared to full-precision methods and achieving threshold accuracy in fewer rounds. Experiments on 100 resource-constrained clients highlight its improved performance and suitability for real-world federated learning on edge devices.

The advent of 5G technology portends a significant increase in Internet of Things (IoT) connectivity, with estimates suggesting over 29 billion devices globally by 2027. This surge underscores the necessity for precise network traffic forecasting in network management. Our study addresses this gap by introducing an Artificial Intelligence (AI)-based forecasting method that amalgamates the Graph Convolutional Network (GCN), Transformer’s encoder, and Temporal Convolutional Network (TCN) components. The GCN, adept at processing network structures, is utilized to discern complex network topologies and spatial dependencies. To encapsulate the temporal dynamics of network traffic, the Transformer encoder is integrated with TCN, enabling simultaneous extraction of both global and local temporal attributes. The performance of our proposed model is rigorously tested on simulated datasets, representing diverse network intensities, derived from GEANT2 and NSFNET frameworks via OMNet++. These simulations focus on one-hop link predictions to evaluate the efficacy of the model. Our findings attest to the model’s robust capability, showcasing its proficiency in accurate network traffic prediction within an IoT milieu dominated by 5G technology.

In case of deploying additional network security equipment in a new location, network service providers face difficulties such as precise management of large number of network security equipment and expensive network operation costs. Accordingly, there is a need for a method for security-aware network service provisioning using the existing network security equipment. In order to solve this problem, there is an existing reinforcement learning-based routing decision method fixed for each node. This method performs repeatedly until a routing decision satisfying end-to-end security constraints is achieved. This generates a disadvantage of longer network service provisioning time. In this paper, we propose security constraints reinforcement learning based routing (SCRR) algorithm that generates routing decisions, which satisfies end-to-end security constraints by giving conditional reward values according to the agent state-action pairs when performing reinforcement learning.

Recently, research on federated clustering has been actively studied to improve the performance of federated learning to solve the non-i.i.d issue. Federated clustering makes clusters with members who has similar characteristics of data which is used as inputs of federated learning, and each cluster trains an artificial intelligence model in a federated manner. However, if distances between members of a cluster configured through federated clustering is long in a network, the overhead related to federated learning becomes larger than expected and it may be lose the network cost benefits of federated learning. In this paper, we propose a DTW(Dynamic Time Warping) based federated clustering and MIP(Mixed Integer Programming)-based edge server placement in order to reduce the network overhead of federated learning caused by federated clustering under non-i.i.d setting.

Recently, with rising attention and widespread awareness of 5G technology, the rapid growth of mobile devices and various network infrastructures and services emerge. As means to provide responsive services and a guaranteed QoS level to individual demands while maintaining resource constraints, it is necessary to consider various factors affecting network service performance and dynamic network provisioning. In this paper, a Reinforcement Learning-based routing algorithm is proposed, which uses the information related to link stability to make routing decisions, called Reinforcement learning-based Routing with Link Stability (RRLS). To evaluate this algorithm, we applied the RRLS algorithm on a dynamic network provisioning framework and compared it to the RRLS algorithm and Dijkstra's algorithm. The result shows that the proposed algorithm performed better than Dijkstra's algorithm and shows that the proposed approach is an appealing solution for dynamic network provisioning routing.

Because of the activation of IoT (Internet of Things) devices due to the rapid development of recent communication technology, network traffic is currently fluctuating and increasing explosively. As existing network resource management policies are not sophisticated enough to cope with network conditions that change constantly, resource utilization can be lowered and costs can be higher. With the recent advances in deep learning techniques, network operators can manage networks intelligently. For the intelligent network, there is a technique which predict the state of network links. However, when the scale of the network increases, overall network management can be complicated. In addition, as the models of link state prediction are affected by the states of adjacent links, it is necessary to consider the spatio-temporal characteristics between links. In this paper, we propose a GCN(Graph Convolutional Neural Network)-GRU(Gated Recurrent Unit) based link state prediction technique. The proposed GCN-GRU model predicts network traffic by considering the spatio-temporal characteristics of each link state such as bandwidth, delay, and packet loss rate. Through extensive experiments on actual network traffic, the proposed GCN-GRU based link state prediction technique has shown to achieve 1.5% lower a mean absolute percentage error (MAPE) compared to a LSTM (Long Short term Memory) based link state prediction technique.

As the number of mobile devices increases and the concept of 5G networks becomes popular, various network infrastructures and services emerge. Also, more users request user-specific network services within limited network resources. Under this complex situation, in order to provide a guaranteed QoS level to users, it requires to consider various factors which affect the performance of network service, and dynamic network provisioning is required. In this paper, we propose a dynamic network provisioning system with AI-enabled path planning, which uses the side channel information such as disaster events, maintenance events, and distribution of users. In this system, we design a user request handler that understands user-specific QoS and spatio-temporal requirements in order to maximize the utilization of a given network resource. Also, this system utilizes side-channel information to optimizing the network provisioning in a realtime manner.

Recently, as the threat of Distributed Denial-of-Service attacks exploiting IoT devices has spread, source-side Denial-of-Service attack detection methods are being studied in order to quickly detect attacks and find their locations. Moreover, to mitigate the limitation of local view of source-side detection, a collaborative attack detection technique is required to share detection results on each source-side network. In this paper, a new collaborative source-side DDoS attack detection method is proposed for detecting DDoS attacks on multiple networks more correctly, by considering the detecting performance on different time zone. The results of individual attack detection on each network are weighted based on detection rate and false positive rate corresponding to the time zone of each network. By gathering the weighted detection results, the proposed method determines whether a DDoS attack happens. Through extensive evaluation with real network traffic data, it is confirmed that the proposed method reduces false positive rate by 35% while maintaining high detection rate.

This paper considers the problem of dynamic evacuation routing in large smart buildings. We investigate a scalable routing approach which not only generates effective routes for evacuees but also quickly updates routes as the disaster status and building conditions could change during the evacuation time. We first design a flexible and scalable evacuation system for large smart buildings with multiple levels of computational support. Given such a system, we develop a novel distributed algorithm for finding effective evacuation routes dynamically by using an LCDT (Length-Capacity-Density-Trustiness) weighted graph model, which is built upon the current disaster information and building conditions. Finally, we propose a caching strategy which expedites dynamic route generation with the current effective route part(s) in order to improve the performance of dynamic evacuation in large buildings. To validate our approach, we test the proposed algorithm with our implementation of an evacuation simulator and compare the results with other approaches. Experimental results show that our approach outperforms other ones in the aspect of the evacuation time reduction and the maximum number of people being evacuated in each time span.

Packet sampling is commonly deployed in all of Intrusion Detection System (IDS) to block the resources consumed from DDoS attack in the network. The IDS usually places sample collector either at distributed points in the network or next to the victim. The sampling methods use the threshold of traffic to detect the attack. It will stop an attack if having a matching with the threshold which is defined in the rule of IDS. We assume that there are lots of suspicious traffics with small volume going to the same destination. The IDS with distributed sampling method cannot detect these traffics. Because of small volume of suspicious traffic, it is not large enough to match the threshold in the rule. Although suspicious traffics have small volume, lots of the traffics generate a large volume at the same destination of victim. To handle this problem, placing the sample collector next to the victim and implementing the destination rule in IDS are proposed, the destination rule can detect multiple traffics that have the same destination IP. However, it still gets the problem that it lets the suspicious traffic going through the network, which generates a large number of unnecessary traffics and causes to the low performance of network. In this paper, we propose a sampling based approach that samples the traffic at the edge gateway in a Software Defined Networking (SDN) based network. It could detect the DDoS attack earlier before reaching to the complex core network, and it could determine which domain the DDoS attack comes from.

With the wide development of network services and the increasing network quality of service (QoS) requirements, dynamic network provisioning, which offers a smart way to segment the network to support particular services, becomes an important problem. Past studies of dynamic network provisioning mostly consider dynamic bandwidth and delay assignment in order to maximize the system throughput and minimize communication overhead. These studies leveraged Software Defined Network (SDN) and Network Function Virtualization (NFV) to create dedicated end-to-end virtual networks which are called network slices. But, these works are lack of consideration of provisioning a dynamic network based on request locations where users would use the network to deploy their services. In this paper, a method for generating a network slice dynamically based on requested locations of users is proposed to support the location-aware dynamic network provisioning. The requested locations are obtained by users through a web interface, then the requested locations are encapsulated and sent to SDN controller which knows the geographical location of SDN switches. Then SDN controller sets up a location-aware network slice. To provide a viability of the proposed approach, we implemented the web interface and the protocol to share the requested locations.

The emerging research of SDN and NFV have been promising to provide flexibility in network provisioning based on service requirement. However, current network provision methods of SDN and NFV do not have a formalized model of describing dynamic requirement of the services. The usage of audio/video services may be the example of the services with dynamic requirement. The orchestrator of SDN and NFV needs to understand the service requirement in order to provision the network dynamically and systemically. Hence, innovative service abstraction is needed to model the service requirements. This paper describes our initial effort to design a service abstraction model for enhancing network provisioning in future network. We consider audio/video services as the target service of our initial service abstraction model and develop the XML based service abstraction model. Moreover, we present a communication mechanism to submit and deploy the XML-based service abstraction model to the orchestrator.

Recently, OpenFlow and FlowVisor are the most promising architecture components of the software-defined networks (SDNs). The evolution of these components brought the revolution of utilizing network elements for the shared and virtual network infrastructure, however from the aspects of production networks, it still lacks of some key component. For example, policy-based services provisioning for user-defined requirement is an indispensable part. Hence, innovative aggregation management design for virtual network management system framework must be developed and tested. This paper describes an initial design to user-defined network infrastructure in order to make it capable of running policy-based engineering related experiments. We developed the framework regarding the real use cases upon the shared and virtual network in KREONET infrastructure. Moreover, we also presented perflow-user-defined specifications which include max rate limiter and dynamic priority assignment strategy to enhance the validation performance of our system for production networks.

To disseminate messages from a single source to a large number of targeted receivers, a natural approach is the tree-based application layer multicast (ALM). In time-constrained flash dissemination scenarios, e.g. earthquake early warning, where time is of the essence, the reliable extensions of the tree-based ALM using ack-based failure recovery protocols cannot support reliable dissemination in the timeframe needed. In this paper, we propose FaReCast which exploits path diversity, i.e., exploit the use of multiple data paths, to achieve fast and reliable data dissemination. First, we design a forest-based M2M (Multiple parents-To-Multiple children) ALM structure where every node has multiple children and multiple parents. The intuition is to enable lower dissemination latency through multiple children, while enabling higher reliability through multiple parents. In order to maintain the M2M ALM structure in a scalable and reliable manner, we develop a DHT-based Distributed Configuration Manager. Second, we design multidirectional multicasting algorithms that effectively utilize the multiple data paths in the M2M ALM structure. A key aspect of our reliable dissemination mechanism is that nodes, in addition to communicating the data to children, also selectively disseminate the data to parents and siblings. As compared to trees using traditional multicasting algorithm, we observe an 80 percent improvement in reliability under 20 percent of failed nodes with no significant increase in latency for over 99 percent of the nodes. Moreover, we notice that FaReCast can reduce the network overhead more than 50 percent by tuning the M2M structure, as compared to the other reliable ALM based disseminations.

Anonymity is one of the main virtues of the Internet, as it protects privacy and enables users to express opinions more freely. However, anonymity hinders the assessment of the veracity of assertions that online users make about their identity attributes, such as age or profession. We propose FaceTrust, a system that uses online social networks to provide lightweight identity credentials while preserving a user’s anonymity. FaceTrust employs a “game with a purpose” design to elicit the opinions of the friends of a user about the user’s self-claimed identity attributes, and uses attack-resistant trust inference to assign veracity scores to identity attribute assertions. FaceTrust provides credentials, which a user can use to corroborate his assertions. We evaluate our proposal using a live Facebook deployment and simulations on a crawled social graph. The results show that our veracity scores are strongly correlated with the ground truth, even when dishonest users make up a large fraction of the social network and employ the Sybil attack.

In this paper, we explore a new form of dissemination that arises in distributed, mission-critical applications called Flash Dissemination. This involves the rapid dissemination of rich information to a large number of recipients in a very short period of time. A key characteristic of Flash Dissemination is its unpredictability (e.g., natural hazards), but when invoked it must harness all possible resources to ensure timely delivery of information. Additionally, it must scale to a large number of recipients and perform efficiently in highly heterogeneous (data, network) and failure prone environments. We investigate a peer-based approach based on the simple principle of transferring dissemination load to information receivers using foundations from broadcast networks, gossip theory, and random networks. Gossip-based protocols are well known for being stateless, scalable, and fault-tolerant; however, their performance degrades as content size increases, because of the propagation of redundant gossip messages. In this paper, we propose Concurrent Random Expanding Walkers (CREW), a smart gossip protocol designed to maximize the speed of dissemination by transmitting data only as needed, and by exploiting both intra- and internode concurrency. CREW is designed to support both content and network heterogeneity and deal with transmission failures without sacrificing dissemination speed. We implemented CREW on top of a scalable middleware environment that allows for deployment across several platforms and developed optimizations without compromising on the stateless nature of CREW. We evaluated CREW empirically and compared it to optimized implementations of popular gossip and peer-based systems. Our experiments show that CREW significantly outperforms both traditional gossip and current large content dissemination systems while sustaining its performance in the presence of network errors.

The eventual goal of any notification system is to deliver appropriate messages to all relevant recipients with very high reliability in a timely manner. In particular, we focus on notification in extreme situations (e.g. disasters) where geographically correlated failures hinder the ability to reach recipients inside the corresponding failed region. In this paper, we present GSFord, a reliable geo-social notification system that is aware of (a) the geographies in which the message needs to be disseminated and (b) the social network characteristics of the intended recipient, in order to maximize/increase the coverage and reliability. GSFord builds robust geo-aware P2P overlays to provide efficient location-based message delivery and reliable storage of geo-social information of recipients. When an event occurs, GSFord is able to efficiently deliver the message to recipients who are either (a) located in the event area or (b) socially correlated to the event (e.g. relatives/friends of those who are impacted by an event). Furthermore, GSFord leverages the geo-social information to trigger a social diffusion process, which operates through out-of band channels such as phone calls and human contacts, in order to reach recipients which are isolated in the failed region. Through extensive evaluations, we show that GSFord is reliable, the social diffusion process enhanced by GSFord reaches up to 99.9\% of desired recipients even under massive geographically correlated regional failures. We also show that GSFord is efficient even under skewed distribution of user populations.

We propose SocialFilter, a trust-aware collaborative spam mitigation system. Our proposal enables nodes with no email classification functionality to query the network on whether a host is a spammer. It employs Sybil-resilient trust inference to weigh the reports concerning spamming hosts that collaborating spam-detecting nodes (reporters) submit to the system. It weighs the spam reports according to the trustworthiness of their reporters to derive a measure of the system's belief that a host is a spammer. SocialFilter is the first collaborative unwanted traffic mitigation system that assesses the trustworthiness of spam reporters by both auditing their reports and by leveraging the social network of the reporters' administrators. The design and evaluation of our proposal offers us the following lessons: a) it is plausible to introduce Sybil-resilient Online-Social-Network-based trust inference mechanisms to improve the reliability and the attack-resistance of collaborative spam mitigation; b) using social links to obtain the trustworthiness of reports concerning spammers can result in comparable spam-blocking effectiveness with approaches that use social links to rate-limit spam (e.g., Ostra); c) unlike Ostra, in the absence of reports that incriminate benign email senders, SocialFilter yields no false positives.

This paper addresses reliability of data dissemination applications when there are severe disruptions to the underlying physical infrastructure. Such massive simultaneous physical failures can happen during the geographical events such as natural disasters (earthquakes, floods, tornados) or sudden power outages - infrastructure failures in these cases are geographically correlated. In particular, we focus on overlay based data dissemination mechanisms and explore their ability to tolerate such large geographically correlated failures. Due to the tight correlation between multiple overlay links and a single physical link, a few physical failures may affect lots of overlay links. To enable reliable dissemination under such conditions, we propose overlay network construction methods that incorporate proximity-aware neighbor selection methods to improve the performance of the overlay data dissemination, to the extent possible, in terms of reliability and latency. In this approach, the overlay nodes select neighbors which are most likely distinct in presence of a geographical failure; we show how an overlay structure constructed using our proximity-aware neighbor selection techniques can disseminate data to over 80% of reachable end clients without any significant additional latency under various geographical failure conditions.

With the recent explosion in usage of the World Wide Web, the problem of caching Web objects has gained considerable importance. The performance of these Web caches is highly affected by the replacement algorithm. Today, many replacement algorithms have been proposed for Web caching and these algorithms use the on-line fashion parameters. Recent studies suggest that the correlation between the on-line fashion parameters and the object popularity in the proxy cache are weakening due to the efficient client caches. We suggest a new algorithm, called Least Popularity Per Byte Replacement (LPPB-R). We use the popularity value as the long-term measurements of request frequency to make up for the weak point of the previous algorithms in the proxy cache and vary the popularity value by changing the impact factor easily to adjust the peformance to needs of the proxy cache. We examine the performance of this and other replacement algorithms via trace driven simulation.