Fifth-generation (5G) cellular networks promise unprecedented connectivity through ultra-low latency and high-speed mobile broadband, driving the need for intelligent slice placement strategies in Open Radio Access Network (O-RAN) architectures. O-RAN promotes openness and vendor interoperability, but existing Machine Learning (ML)-based embedding solutions often prioritize performance metrics such as delay and availability while neglecting energy efficiency. To address this gap, we propose Crown, a Reinforcement Learning (RL)-based service placement framework that jointly optimizes Service Level Agreement (SLA) compliance and power consumption. Crown extends a traditional Deep Q-Network (DQN) by integrating cross-attention layers to model complex dependencies between virtualized O-RAN functions and heterogeneous physical servers, enabling more informed placement decisions. We evaluate Crown in a simulated O-RAN environment and compare it against state-of-the-art RL approaches and heuristic baselines. Results demonstrate that Crown reduces power consumption by 57% compared to a fixed deployment and by 15% relative to a DQN without cross‑attention, while meeting stringent latency and bandwidth requirements through action masking and achieving high slice admission rates and low deployment cost via its cost‑aware reward design. Furthermore, we measure the inference time, showing that the attention-enhanced RL design remains practical for large-scale deployments.

Teaching a machine to accurately identify human activities from sensor data poses a significant challenge, which is further compounded by considerations of data privacy, resource costs, and responsiveness, particularly within the constraints of devices like smartphones. While current solutions efficiently identify activities, trained models are barely portable in scenarios composed of diverse activities and limited battery life devices, such as smartphones. This paper introduces Transferable and Copyright-Preserving Human Activity Recognition (TCP-HAR), a mobile-based HAR system that integrates digital watermarking, Federated Learning (FL), Transfer Learning (TL), and compression techniques to provide efficient human activity recognition while providing copyright protection of deep neural network models over Android smartphones. Our solution optimizes the utilization of FL, TL, and their combination (FTL) by extensively testing standalone TL models in offline contexts and comparing these results with FL across a network of mobile devices. Our findings highlight the benefits of TCP-HAR for mobile environments in terms of accuracy, F1-score, and training time. In addition, our proposed watermarking mechanism is robust yet computationally efficient, ensuring ownership verification without compromising the scalability of the TFL process.

Network traffic telemetry plays a crucial role in the management of modern networks. Estimation of the network traffic matrix is a widely recognized problem whose solutions can span a diverse set of applications. Current approaches to traffic matrix inference through statistical methods often rely on assumptions about the matrix structure, which may be invalid in certain scenarios. Data-driven methods, instead, often use detailed information about the network topology that may be unavailable or impractical to collect. To overcome these challenges, we propose a super-resolution technique for traffic matrix inference that leverages coarser measurements to predict fine-grained network traffic. Furthermore, we devise a distributed learning procedure and adapt our model to scenarios of partial network visibility. Our experiments on real network traces demonstrate that the proposed approach can infer fine-grained network traffic with high precision. Moreover, we prove that our distributed approach improves the inference accuracy with respect to its centralized counterpart, significantly lowering the training time, even in scenarios with partial network knowledge.

Autonomous exploration of complex, unknown environments is a cutting-edge task not entirely solved by the scientific community. When an agent needs to explore a maze without any a priori information about the environment, the lack of proper destinations and explicit task objectives make traditional navigation policies inappropriate. While the literature presents some sporadic deterministic systems able to face the tasks, learning approaches still need an adequate investigation which could prove them to be more suitable and versatile for this purpose. In this paper, we present MARS, a path planner that exploits swarms of robots to optimize the exploration of complex unknown environments, such as mazes. To make the solution scalable, the proposed method exploits two cooperating modules: local and global planners. The local planner is modeled as a Markov Decision Process (MDP) and trained as a Reinforcement Learning (RL) multi-agent system. Each agent has access to image representations of a section of the global map, always centered in the robot reference frame, and decides the next navigation goal to complete the local exploration. The global planner is a deterministic system that recovers the navigation when a local solution is unavailable. The robots share the explored section with peers when they meet in a rendez-vous. We compared our approach to a single deterministic agent, a single RL agent and a close-to-optimal deterministic approach which deploys five greedy agents. The simulation results demonstrate MARS' efficiency, reaching near-optimal levels in significantly less time.

Network traffic is now largely encrypted. Yet analysis of side-channel features-packet sizes, timings, and directions-can still reveal patterns about encrypted flows. Recent machine learning (ML) techniques have made such traffic analysis more powerful, and they can be applied both offensively (e.g., inference attacks) and defensively (e.g., intrusion detection). This raises the need for protections that keep pace with MLenabled capabilities without exacerbating resource overhead and reaction delays. My research explores new opportunities, such as application-agnostic defenses, offered by data-plane programmability (e.g., eBPF/XDP at hosts and P4 in switches) to reshape observable traffic patterns and fast feature extraction for advanced detection mechanisms. My PhD also focuses on designing and prototyping the combination of ML together with programmable data planes to both mitigate traffic analysis and harness it for defense, while clarifying the trade-offs between privacy, performance, and deployability.

As our digital presence expands, safeguarding private data and preserving online privacy becomes paramount. Thus, motivating the development of secure DNS systems, such as DNS over TLS or HTTPS. The vulnerability of these protocols against privacy attacks has led to the development of the Oblivious DNS-over-HTTPS (ODoH) protocol. Nevertheless, the extent of ODoH’s effectiveness in protecting clients’ privacy is still unknown. This study investigates ODoH resiliency against website fingerprinting attacks in the open-world setting. We deploy an ODoH testbed on GENI for data collection and employ deep learning techniques such as ensemble learning for data analysis. Our findings reveal that a passive adversary can identify targeted websites using ODoH traces with an accuracy of 94%. Additionally, we analyze the impact of various factors, including clients’ locations, available resolvers, and time stability, on the attack’s success. Finally, we prototype a mitigation strategy and demonstrate its effectiveness in safeguarding clients privacy.

Effective congestion control is critical for wireless networks, where rapidly varying channel conditions and diverse traffic demands can severely degrade performance. Traditional congestion control algorithms rely on static heuristics that are often ill-suited for dynamic wireless environments. In this paper, we introduce Flecto, a Reinforcement Learning (RL)-based congestion control solution integrated into the QUIC protocol that, leveraging cross-layer metrics, including Signal-to-Noise Ratio, Block Error Rates, and Round-Trip Time measurements, can take decisions using a comprehensive view of network conditions. We implemented Flecto on a 5G testbed using OpenAirInterface and ETTUS USRP B210 radios, showing how it adapts transmission rates in real-time to maximize throughput and minimize latency while maintaining stability. Experimental results show that Flecto achieves an average throughput of 4539.5 KB/s approximately 6% higher both than Cubic (4267.2 KB/s) and New Reno (2674.1 KB/s) while reducing the average Round-Trip Time to 21.8 ms, significantly lower than Cubic’s 27.6 ms and New Reno’s 174.9 ms. These performance gains underscore the promise of integrating RL with cross-layer feedback for adaptive, efficient congestion control in next-generation wireless networks. Moreover, the modular design of Flecto facilitates its extension to other transport protocols and multi-user scheduling frameworks, paving the way for broader adoption in future wireless systems.

Microbursts - microsecond-scale congestion events - are a major cause of packet loss and performance degradation in modern datacenter networks. While packet deflection techniques can help manage microbursts, current implementations lead to excessive packet reordering, exacerbated congestion under high load, and head-of-line blocking in switch buffers. In this paper, we design and implement RobinHood, a novel in-network burst-tolerant protocol. At its core, the protocols mechanisms and policies are based on work-stealing, a technique originally designed to reduce job completion times in operating systems. Through extensive trace-driven simulations on leaf-spine and fattree topologies, we show that RobinHood improves flow completion times up to 22% over Equal-Cost Multi-Path (ECMP), and up to 7% over recent solutions, DIBS and Vertigo, under high load scenarios.

The advent of Large Language Models (LLMs) is progressively transforming how complex tasks across various domains can be automated, with a notable potential impact on cloud computing operations. In this domain, LLMs might be used, for example, to configure Kubernetes (K8s) clusters via the generation of manifest files – structured configuration files defining the containerized environment. However, despite the considerable advances in LLMs’ text generation, this task conceals several challenges that prevent operators from achieving a fully automated process. In this paper, we present the current trends in solving these gaps, quantitatively evaluate the accuracy of LLM-based approaches to generate K8s manifests starting from human intents, and discuss open challenges that make benchmarking and automation still complex. Experiments over three open-source LLMs demonstrate how intent-based K8s manifest generation can be effectively achieved through model fine-tuning, but also what open issues remain and must be addressed prior to having an autonomous and self-healing K8s infrastructure managed via Agentic AI.

In recent decades, applications in environmental sustainability, education, and housekeeping have become increasingly distributed and sophisticated, leveraging a wide range of devices to perform complex tasks. While a large number of agents can reduce computation time, managing these distributed systems presents significant challenges due to resource constraints such as power consumption and storage. To address this, the literature has explored various model compression techniques, such as pruning, to optimize performance in distributed environments. In this paper, we propose DFL-Trim, a solution for trimming models in Decentralized Federated Learning (FL) that meets network constraints while maintaining satisfactory performance. We demonstrate how pruning can be implemented in decentralized settings, analyze its effect on bandwidth usage, and discuss the trade-offs between compression and model accuracy.

Network topology verification in Software-Defined Networks (SDN) poses a significant challenge, as vulnerabilities can allow attackers to deceive the controller and manipulate the data plane into incorrect topologies, thereby endangering the entire network's security. Current solutions fail to guarantee both security and efficiency in the verification process, often resulting in damaging user traffic. With the aim of solving joint objectives, in this paper, we introduce PathSafe, a novel tool constructed on top of the existing controller frameworks designed for secure path verification in SDN environments. It enables the verification of all available paths between two points in the network and ensures a secure process. Our approach requires a data plane component for real-time packet monitoring at line speed and a control plane verification step. Our research demonstrates that PathSafe effectively mitigates security risks in compromised switches and host scenarios. Alongside a theoretical exploration of this challenge, we present a proof of concept implemented in P4, a common language for programmable data planes. Results obtained in Mininet underscore the practical applicability of PathSafe that, compared to alternatives, reduces overhead in the verification process while maintaining a limited execution time.

Service orchestrators such as Kubernetes are widely employed to automate the handling and scheduling of workloads, which involves determining the most suitable physical node on which to start a new task. The expanding application of Machine Learning (ML) algorithms, and in particular Reinforcement Learning (RL), opens up new development opportunities to make runtime decisions that can account for multiple metrics and varying network conditions. However, current RL-based solutions are unable to fit the growing complexity of distributed applications and infrastructure, characterized by a more heterogeneous resource continuum and the increasing need to minimize energy consumption while satisfying tasks' requirements. To fill this gap, we propose RL-ICE as an innovative scheduler that can work in such a cloud continuum by leveraging a multi-cluster and hierarchical RL to satisfy both user Quality of Experience (QoE) metrics and tenant's costs. We test RL-ICE in a simulated large-scale environment and in a real-world Kubernetes setup. In both scenarios, our solution effectively balances user-perceived latency, energy consumption, and deployment costs. Additionally, RL-ICE can dynamically respond to network failures by migrating microservices to maintain efficient management of resources.

Cloud gaming represents a rapidly growing segment in the entertainment industry, allowing users to stream and interact with high-quality games over the Internet. However, the problem of maintaining a seamless gaming experience is inherent to minimizing user-perceived latency. In this paper, we present CLoud Application lAtency Prediction (CLAAP), a novel solution that, to tolerate challenged network conditions in gaming, predicts such latency via a Machine Learning (ML) model and forecasts future network evolution. The model, trained over diverse network conditions and gaming scenarios, can then update its parameters via a concept drift detection algorithm that suggests a re-training action, reducing the prediction error up to 21% with minimal overhead. We then integrate this network metrics predictor into a game state prediction to further tolerate network latency spikes even from the user perspective, who can continue playing even in adversarial conditions without session interruptions. The results suggest the potential of advanced predictive analytics in mitigating latency issues, thereby setting the stage for more responsive and immersive cloud gaming services.

Federated Learning has revolutionized the way in which mobile devices and IoT can share common knowledge in data analytics. However, some challenges arise when dealing with heterogeneous and challenged networks, especially in gradient synchronization. For example, some clients (referred to as stragglers) may take much longer to report their output than other nodes. Current solutions addressing the straggling problems either propose a distributed coordination (but introduce new synchronization issues) or deadline-based approaches to discard clients after a fixed deadline (but introduce the problem of determining a suitable deadline). To this end, we propose to set a dynamic deadline in which the central server selects the best IoT nodes via an online learning approach based on predicting the response time of each client. Moreover, to further mitigate synchronization and scalability issues, we also consider a hierarchical approach in which clients send model parameters to intermediate aggregation edge servers. Our results demonstrate that this approach can lower network overhead by 78% compared to the widely adopted FedAvg and 49% to the best alternative. At the same time, the model accuracy is preserved, and the training time in challenged networks is reduced by 52% w.r.t. FedAvg and 32% w.r.t. recent solutions.

Traditional network management requires manual coding and expertise, making it challenging for non-specialists and experts to handle increasing devices and applications. In response, Intent-Based Networking (IBN) has been proposed to simplify network operations by allowing users to express in natural language the program objective (or intent), which is then translated into device-specific configurations. The emergence of Large Language Models (LLMs) has boosted the capabilities to interpret human intents, with recent IBN solutions embracing LLMs for a more accurate translation. However, while these solutions excel at intent comprehension, they lack a complete pipeline that can receive user intents and deploy network programs across devices programmed in multiple languages. In this paper, we present LLNeT, our IBN solution that, within the context of Software-Defined Networking (SDN), can translate seamlessly intent-to-program. First, leveraging LLMs, we convert network intents into an intermediate representation by extracting key information; then, using this output, the system can tailor the network code for any topology using the specific language calls. At the same time, we address the challenge of a more sustainable IBN approach to reduce its energy consumption, and we experience how even a Small Language Model (SLM) can efficiently help LLNeT for input translation. Results across multiple use cases demonstrated how our solution can guarantee adequate translation accuracy while reducing operator expenses compared to other LLM-based approaches.

AI/ML has enhanced computer networking, aiding administrators in decision-making and automating tasks for optimized performance. Despite such advances in network automation, there remains limited trust in these uninterpretable models due to their inherent complexity. To this aim, eXplainable AI (XAI) has emerged as a critical area to demystify (deep) neural network models and to provide more transparent decision-making processes. While other fields have embraced XAI more prominently, the use of these techniques in computer network management remains largely unexplored. In this paper, we shed some light by presenting, an XAI-based approach designed to clarify the opaque nature of data-driven traffic engineering solutions in general, and efficient network telemetry, in particular. It does so by examining the intrinsic behavior of the adopted models, thereby reducing the volume of data needed for effective learning. Our extensive evaluation revealed how our approach not only reduces training time and overhead in network telemetry models but also maintains or improves model accuracy, leading, in turn, to more efficient and clear ML models for network management.

Network measurement and telemetry techniques are central to the management of modern computer networks. Internet traffic matrix estimation is a popular technique employed for network management and telemetry to reconstruct missing information. Existing approaches use statistical methods, which often make impractical assumptions about the structure of the Internet traffic matrix. Data-driven methods, instead, heavily rely on the assumption of full knowledge of network topology data, that may be unavailable or impractical to collect. In this work, we propose ResCue, a deep residual networks technique to infer fine-grained Internet network traffic starting from spatial coarse-grained measurements. To address scenarios with network visibility constraints, we design a federated learning approach for fine-grained traffic prediction with partial network knowledge. Our evaluation across real-world traffic data shows that our proposed approach outperforms existing interpolation techniques and that our federated learning design achieves similar accuracy with respect to its centralized counterpart while requiring only partial knowledge of the network.

Recent advances in Machine Learning (ML) brought several advantages also within computer network management. For programmable data planes, however, it is more challenging to benefit from these advantages, given their limited resource capabilities colliding with the complexity of ML models. In this paper, we propose ART, an attempt to simplify ML-based solutions for routing, so that they can "fit", i.e., be executed, on P4 switches. To provide such model simplification, ART relies on efficient knowledge distillation techniques, converting, in particular, Deep Reinforcement Learning (DRL) models into a simpler Decision Tree (DT). Our evaluation results validate the accuracy of the extracted model and the application of the model logic directly into switches with little impact, paving the way for a more reactive data plane programmability via machine learning integration.

Efficient service provisioning in the Edge-to-Cloud Continuum is of utmost importance for modern applications. While sensible decisions can be taken if enough monitoring data is collected, maintaining continuous telemetry data streams amidst the continuum’s complexity is challenging. This paper introduces CRISP (reConstructing Resource Information for Service Placement), a solution combining data reconstruction and service placement strategies to optimize decisions despite incomplete monitoring data. CRISP utilizes Convolutional Neural Networks and Long Short-Term Memory models for data reconstruction, integrating them with a heuristic algorithm that selects nodes for service component placement. Numerical results demonstrate CRISP’s efficacy in optimizing service provisioning despite missing data, contributing to enhanced resource utilization and service performance in the considered context.

In recent years, containerized deployment models have gained favor across many domain of applications. Kubernetes, the de-facto standard for containers orchestration, can efficiently manage heterogeneous devices, but fails to adapt to possibly stringent requirements, as it only considers computing metrics for scheduling decisions. In addition, the rising prominence of distributed cloud environments, which enable the development of highly available, performant solutions, requires modifications to the default Kubernetes scheduler. To address these challenges, we introduce LAIS, a multi-cluster Kubernetes scheduler optimized for end-to-end latency measurements to enhance user Quality of Experience (QoE). Unlike existing approaches, we define a geographically distributed environment and deploy a solution that satisfies user-specified intents in terms of latency. Depending on user needs, LAIS can either meet a specific latency constraint or schedule pods in the cluster with the lowest latency. After implementing LAIS in a multi-cluster environment, we found it highly effective in accommodating a range of user intents, outperforming the default Kubernetes scheduler in this regard.

A large fraction of recent network management tasks rely on Internet traffic matrices, ranging from planning and troubleshooting to routing and anomaly detection. Despite extensive research efforts over the years, acquiring a comprehensive overview of network traffic remains a difficult and error-prone task. While the literature has mostly proposed increasingly accurate and complex Machine Learning (ML) models to reconstruct missing information, in this paper we propose an alternative approach to further enhance this process: combining the ML model with eXplainable AI (XAI) to analyze the model behavior, detect most significant features, and limit the reconstruction process to such reduced input. With this methodology, not only we simplify the problem, but the entire solution finds greater deployability as the data acquisition phase is also simplified. Numerical results demonstrate that, with our solution on a Convolution Neural Network model, the error during completion can be lowered by 80% for a network telemetry traffic reduction of 75%.

Data center traffic management challenges increase with the complexity and variety of new Internet and Web applications. Efficient network management systems are often needed to thwart delays and minimize failures. In this regard, it seems helpful to identify in advance the different classes of flows that (co)exist in the network, characterizing them into different types based on different latency/bandwidth requirements. In this paper, we propose Howdah, a traffic identification and profiling mechanism that uses Machine Learning and a load-aware forwarding strategy to offer adaptation to different classes of traffic with the support of programmable data planes. With Howdah, the sender and gateway elements inject in-band traffic information obtained by a supervised learning algorithm. When a switch or router receives a packet, it exploits this host-based traffic classification to adapt to a desirable traffic profile, for example, to balance the traffic load. We compare our solution against recent traffic engineering proposals and demonstrate the effectiveness of the cooperation between host traffic classification and P4-based switch forwarding policies, reducing packet transmission time in data center scenarios.

Traffic matrices are used for many network management operations, from planning to repairing. Despite years of research on the topic, their estimation and inference on the Internet are still challenging and error-prone. For example, missing values are unavoidable due to flaws in the measurement systems and possible failure in data collection systems. It is thus helpful for many network operators to recover the missing data from the partial direct measurements. Some existing matrix completion methods do not fully consider network traffic behavior and hidden traffic characteristics, showing the inability to adapt to multiple scenarios. Others instead make assumptions on the matrix structure that may be invalid or impractical, curtailing the applicability. In this paper, we propose Hide & Seek, a novel matrix completion and prediction algorithm based on a combination of generative autoencoders and Hidden Markov Models. We demonstrate with an extensive experimental evaluation on real-world datasets how our algorithm can accurately reconstruct missing values while predicting their short-term evolution.

Years of research on congestion controls have highlighted how end-to-end and in-network protocols might perform poorly in some contexts. Recent advances in data plane network programmability could also bring advantages in transport protocols, enabling mining and processing in-network congestion signals. However, the new machine learning-based congestion control class has only partially used data from the network, favoring a more sophisticated model design but neglecting possibly precious pieces of data. In this paper, we present HINT, an in-band network telemetry architecture designed to provide insights into network congestion to the end-host TCP algorithm during the learning process. In particular, the key idea is to adapt switches’ behavior via P4 and instruct them to insert simple device information, such as processing delay and queue occupancy, directly into transferred packets. Initial experimental results show that this approach comes with a little network overhead but can improve the visibility and, consequently, the accuracy of TCP decisions of the end-host. At the same time, the programmability of both switches and hosts also enables customization of the default behavior as the user’s needs change.

A recent trend dictating evolution of management and orchestration of computer networks is constituted by the softwarization and virtualization of them, which have drastically simplified the deployment and real-time reconfiguration of network functions, allowing them to continuously adapt and to deal with dynamic demands in an automated way. Alongside, recent management and orchestration approaches for softwarized networks employ Artificial Intelligence (AI) and Machine Learning (ML) to further reduce reaction time and improve the accuracy of decisions, where the network operations can be automated to the point of realizing autonomous driving networks. However, while automating operations can improve the overall system (it is acknowledged that 70% of network faults are caused by manual errors), AI/ML methods are not the panaceas, and we are still far from having a fully operating and efficient automated architecture. In this dissertation, we present a novel class of software network solutions that share the goal of enabling intelligent and autonomous computer networks, exploring how to exploit the power of AI/ML to handle the growing complexity of critical systems. We start with a new network management scheme for adaptive routing and autonomous scaling of virtual network resources. Then, acting on the hosts, we propose to adjust the TCP congestion control with a ML-based solution, whose goal is to select the proper congestion window learning from end-to-end features and (when available) network signals. We believe that the proposed solutions, and their combination, can lay the foundation for automated systems that better suit modern edge environments and cellular networks by providing unprecedented flexibility and adaptation to even unseen and unknown network conditions.

Recent medical applications are largely dominated by the application of Machine Learning (ML) models to assist expert decisions, leading to disruptive innovations in radiology, pathology, genomics, and hence modern healthcare systems in general. Despite the profitable usage of AI-based algorithms, these data-driven methods are facing issues such as the scarcity and privacy of user data, as well as the difficulty of institutions exchanging medical information. With insufficient data, ML is prevented from reaching its full potential, which is only possible if the database consists of the full spectrum of possible anatomies, pathologies, and input data types. To solve these issues, Federated Learning (FL) appeared as a valuable approach in the medical field, allowing patient data to stay where it is generated. Since an FL setting allows many clients to collaboratively train a model while keeping training data decentralized, it can protect privacy-sensitive medical data. However, FL is still unable to deliver all its promises and meets the more stringent requirements (e.g., latency, security) of a healthcare system based on multiple Internet of Medical Things (IoMT). For example, although no data are shared among the participants by definition in FL systems, some security risks are still present and can be considered as vulnerabilities from multiple aspects. This paper sheds light upon the emerging deployment of FL, provides a broad overview of current approaches and existing challenges, and outlines several directions of future work that are relevant to solving existing problems in federated healthcare, with a particular focus on security and privacy issues.

Internet traffic matrices are used nowadays for a variety of network management operations, from planning to repairing. Despite years of research on the topic, obtaining a global view of traffic is still challenging and error-prone. Due to flaws in the measurement systems and possible failure in data collection tools, missing values are unavoidable. It is thus helpful for many network operators to recover the missing data from the partial direct measurements. While some existing matrix completion methods allowed this reconstruction, they do not fully consider network traffic behavior and hidden traffic characteristics, showing the inability to adapt to multiple scenarios. Others instead make assumptions about the matrix structure that may be invalid or impractical, curtailing the applicability. In this paper, we propose Hide & Seek, a novel matrix completion and prediction algorithm based on a combination of generative autoencoders and Hidden Markov Models. After an extensive experimental evaluation based on both real-world datasets and on a testbed, we demonstrated how our algorithm can accurately reconstruct missing values while also predicting their short-term evolution.

Despite years of research on transport protocols, the tussle between in-network and end-to-end congestion control has not been solved. This debate is due to the variance of conditions and assumptions in different network scenarios, e.g., cellular versus data center networks. Recently, the community has proposed a few transport protocols driven by machine learning, nonetheless limited to end-to-end approaches. In this paper, we present Owl, a transport protocol based on reinforcement learning, whose goal is to select the proper congestion window learning from end-to-end features and network signals, when available. We show that our solution converges to a fair resource allocation after the learning overhead. Our kernel implementation, deployed over emulated and large scale virtual network testbeds, outperforms all benchmark solutions based on end-to-end or in-network congestion control.

The challenges of managing datacenter traffic increase with the complexity and variety of new Internet and Web applications. Efficient network management systems are often required to thwart delays and minimize failures. In this regard, it appears helpful to identify in advance the different classes of flows that (co)exist in the network, characterizing them into different types according to the different latency/bandwidth requirements. In this paper, we propose Howdah, a traffic identification and profiling mechanism that uses Machine Learning and a congestion-aware forwarding strategy to offer adaptation to different traffic classes with the support of programmable data-planes. With Howdah, sender and gateway elements inject in-band traffic information obtained using supervised learning. When a switch or a router receives a packet, it exploits such host-based traffic classification to adapt to a desirable traffic profile, for example, balancing the load. We compare our solutions against recent traffic engineering solutions and show the efficacy of cooperation between host traffic classification and P4-based switch forwarding policies, reducing packet transmission time in datacenter scenarios.

Translating high-level policies to lower-level network rules is one of the main goals of control or data plane network programmability. To further abstract requirements and propel automation in networking, several industries have proposed the paradigm of “network intent”. However, the translation from intents to low-level policies is considered critical to program data planes and other network elements, especially when dealing with P4-enabled switches. In this paper, we present NLP4, an architecture that helps translate intents, in the form of human language, into data-plane programs, in the form of P4 rules. In particular, NLP4 uses Natural Language Processing (NLP) techniques to translate high-level human-language intents, a MultiLayer Perceptron (MLP) model for processing the NLP output and converting it into mid-level policy. An API then uses this information, which separates the intent from the network to generate commands readable by P4-enabled switches. Our initial prototype on a network emulator validates our architecture for a specific case: load profiling, demonstrating how even users with limited P4 expertise may customize their networks by merely specifying intents.

The Internet of Things (IoT), coupled with the edge computing paradigm, is enabling several pervasive networked applications with stringent real-time requirements, such as telemedicine and haptic telecommunications. Recent advances in network virtualization and artificial intelligence are helping solve network latency and capacity problems, learning from several states of the network stack. However, despite such advances, a network architecture able to meet the demands of next-generation networked applications with stringent real-time requirements still has untackled challenges. In this paper, we argue that only using network (or transport) layer information to predict traffic evolution and other network states may be insufficient, and a more holistic approach that considers predictions of application-layer states is needed to repair the inefficiencies of the TCP/IP architecture. Based on this intuition, we present the design and implementation of Reparo. At its core, the design of our solution is based on the detection of a packet loss and its restoration using a Hidden Markov Model (HMM) empowered with adversarial autoencoders. In our evaluation, we considered a telemedicine use case, specifically a telepathology session, in which a microscope is controlled remotely in real-time to assess histological imagery. Our results confirm that the use of adversarial autoencoders enhances the accuracy of the prediction method satisfying our telemedicine application’s requirements with a notable improvement in terms of throughput and latency perceived by the user.

Edge computing technologies have improved delays and privacy of several applications, including in medical imaging and eHealth. In this paper, we consider ultrasound technology and echocardiology (echo) and empower it with edge computing. Despite the many advances that ultrasound technology has seen recently, e.g., it is possible to perform echo scans using wireless ultrasound probes, the use of Artificial Intelligence (AI) techniques is becoming a necessity, for faster and more accurate echo diagnosis (not limited to heart diseases). While a few proprietary solutions exist that embed AI within echo devices, none of them uses resource-intensive tasks on handheld devices, and none of them is open-source. To this end, we propose EdgeEcho, an architecture that captures ultrasound data originated from handheld ultrasound probes and tags it using semantic segmentation performed on edge cloud. Our prototype focuses on optimizing the management of edge resources to address the specific requirements of echocardiology and the challenges of serving AI algorithms responsively. As a use case, we focus on a ventricular volume detection operation. Our performance evaluation results show that EdgeEcho can support multiple parallel medical video processing streaming sessions for continuing medical education, demonstrating a promising edge computing application with life-saving potential.

The soaring complexity of networks has led to more and more complex methods to manage and orchestrate efficiently the multitude of network environments. Several solutions exist, such as OpenFlow, NetConf, P4, DPDK, etc., that allow net-work programmability at both control and data plane level, driving innovation in many focused high-performance networked applications. However, with the increase of strict requirements in critical applications, also the networking architecture and its operations should be redesigned. In particular, recent advances in machine learning have opened new opportunities to the automation of network management, exploiting existing advances in software-defined infrastructures. We argue that the design of effective data-driven network management solutions needs to collect, merge, and process states from both data and control planes. This paper sheds light upon the benefits of utilizing such an approach to support feature extraction and data collection for network automation.

Years of research on transport protocols have not solved the tussle between in-network and end-to-end congestion control. This debate is due to the variance of conditions and assumptions in different network scenarios, e.g., cellular versus data center networks. Recently, the community has proposed a few transport protocols driven by machine learning, nonetheless limited to end-to-end approaches.In this paper, we present Owl, a transport protocol based on reinforcement learning, whose goal is to select the proper congestion window learning from end-to-end features and network signals, when available. We show that our solution converges to a fair resource allocation after the learning overhead. Our kernel implementation, deployed over emulated and large scale virtual network testbeds, outperforms all benchmark solutions based on end-to-end or in-network congestion control.

The abiding attempt of automation has also permeated the networks, with the ability to measure, analyze, and control themselves in an automated manner, by reacting to changes in the environment (e.g., demand). When provided with these features, networks are often labeled as “self-driving” or “autonomous”. In this regard, the provision and orchestration of physical or virtual resources are crucial for both Quality of Service (QoS) guarantees and cost management in the edge/cloud computing environment. To effectively manage the lifecycle of these resources, an auto-scaling mechanism is essential. However, traditional threshold-based and recent Machine Learning (ML)-based policies are often unable to address the soaring complexity of networks due to their centralized approach. By relying on multi-agent reinforcement learning, we propose Mystique, a solution that learns from the load on links to establish the minimal set of active network resources. As traffic demands ebb and flow, our adaptive and self-driving solution can scale up and down and also react to failures in a fully automated, flexible, and efficient manner. Our results demonstrate that the presented solution can reduce network energy consumption while providing an adequate service level, outperforming other benchmark auto-scaling approaches.

The abiding attempt of automation has also pervaded computer networks, with the ability to measure, analyze, and control themselves in an automated manner, by reacting to changes in the environment (e.g., demand) while exploiting existing flexibilities. When provided with these features, networks are often referred to as "self-driving". Network virtualization and machine learning are the drivers. In this regard, the provision and orchestration of physical or virtual resources are crucial for both Quality of Service guarantees and cost management in the edge/cloud computing ecosystem. Auto-scaling mechanisms are hence essential to effectively manage the lifecycle of network resources. In this poster, we propose Relevant, a distributed reinforcement learning approach to enable distributed automation for network orchestrators. Our solution aims at solving the congestion control problem within Software-Defined Network infrastructures, while being mindful of the energy consumption, helping resources to scale up and down as traffic demands fluctuate and energy optimization opportunities arise.

The edge computing paradigm allows computationally intensive tasks to be offloaded from small devices to nearby (more) powerful servers, via an edge network. The intersection between such edge computing paradigm and Machine Learning (ML), in general, and deep learning in particular, has brought to light several advantages for network operators: from automating management tasks, to gain additional insights on their networks. Most of the existing approaches that use ML to drive routing and traffic control decisions are valuable but rarely focus on challenged networks, that are characterized by continually varying network conditions and the high volume of traffic generated by edge devices. In particular, recently proposed distributed ML-based architectures require either a long synchronization phase or a training phase that is unsustainable for challenged networks. In this paper, we fill this knowledge gap with Blaster, a federated architecture for routing packets within a distributed edge network, to improve the application's performance and allow scalability of data-intensive applications. We also propose a novel path selection model that uses Long Short Term Memory (LSTM) to predict the optimal route. Finally, we present some initial results obtained by testing our approach via simulations and with a prototype deployed over the GENI testbed. By leveraging a Federated Learning (FL) model, our approach shows that we can optimize the communication between SDN controllers, preserving bandwidth for the data traffic.

Distributed systems are extremely difficult to design and implement correctly because they must handle both system correctness and device failures. Most of the work focuses on the first aspect, and in particular, on the correctness of security and network configuration. The large demand for availability and reliability for critical services is actually pushing new architectures that tolerate faults, but a-priori analysis of redundancy and recovery features is still limited. To this end, we present a framework to design and formally verify the persistence of network properties, even in case of failures. The solution considers both nodes and links failure, and it is based on a formal model that takes both network topology and network device configurations into account. In contrast, most of the existing approaches only consider network topology. By analyzing the formal model, the framework can check whether the specified network services are still available after failures, and in case of success, it outputs a possible configuration of the devices to be used for automatic recovery.

Telepathology aims to replace the pathology operations performed on-site, but current systems are limited by their prohibitive cost, or by the adopted underlying technologies. In this work, we contribute to overcoming these limitations by bringing the recent advances of edge computing to reduce latency and increase local computation abilities to the pathology ecosystem. In particular, this paper presents LiveMicro, a system whose benefit is twofold: on one hand, it enables edge computing driven digital pathology computations, such as data-driven image processing on a live capture of the microscope. On the other hand, our system allows remote pathologists to diagnosis in collaboration in a single virtual microscope session, facilitating continuous medical education and remote consultation, crucial for under-served and remote hospital or private practice. Our results show the benefits and the principles underpinning our solution, with particular emphasis on how the pathologists interact with our application. Additionally, we developed simple yet effective diagnosis-aided algorithms to demonstrate the practicality of our approach.

The demand of low latency applications has fostered interest in edge computing, a recent paradigm in which data is processed locally, at the edge of the network. The challenge of delivering services with low-latency and high bandwidth requirements has seen the flourishing of Software-Defined Networking (SDN) solutions that utilize ad-hoc data-driven statistical learning solutions to dynamically steer edge computing resources. In this paper, we propose RoPE, an architecture that adapts the routing strategy of the underlying edge network based on future available bandwidth. The bandwidth prediction method is a policy that we adjust dynamically based on the required time-to-solution and on the available data. An SDN controller keeps track of past link loads and takes a new route if the current path is predicted to be congested. We tested RoPE on different use case applications comparing different well-known prediction policies. Our evaluation results demonstrate that our adaptive solution outperforms other ad-hoc routing solutions and edge-based applications, in turn, benefit from adaptive routing, as long as the prediction is accurate and easy to obtain.

Software-Defined Networking (SDN) is a paradigm that enables easier network programmability based on separation between network control plane and data plane. Network Function Virtualization (NFV) is another recent technology that has enabled design, deploy, and management of softwarized networking services. The vast majority of SDN and NFV based architectures, whether they use Virtual machines (VMs) or Lightweight Virtual Machines (LVMs), are designed to program forwarding, probably the most fundamental among all network mechanisms. In this paper instead we demonstrated that there are other (as important) networking mechanisms that need programmability. In particular, we designed, implemented and extensively tested an architecture that enables policy-programmability of (live) migration of LVMs. Migration is used for maintenance, load balancing, or as a security mechanism in what is called Moving Target Defence (a virtual host migrates to hide from an attacker). Our architecture is based on Docker and it is implemented within a Software-Defined Infrastructure. Migration mechanism can be set easily by means of configuration file, to make a novel policy-based architecture. We evaluated the performance of our system in several scenarios, over a local Mininet-based testbed. We analyzed the tradeoff between several Load Balancing policies as well as several Moving Target Defense solutions inspired by network coding.