With the massive amount of data generated from mobile devices and the increase of computing power of edge devices, the paradigm of Federated Learning has attracted great momentum. In federated learning, distributed and heterogeneous nodes collaborate to learn model parameters. However, while providing benefits such as privacy by design and reduced latency, the heterogeneous network present challenges to the synchronisation methods, or barrier control methods, used in training, regarding system progress and model convergence etc. The design of these barrier mechanisms is critical for the performance and scalability of federated learning systems. We propose a new barrier control technique called Probabilistic Synchronous Parallel (PSP). In contrast to existing mechanisms, it introduces a sampling primitive that composes with existing barrier control mechanisms to produce a family of mechanisms with improved convergence speed and scalability. Our proposal is supported with a convergence analysis of PSP-based SGD algorithm. In practice, we also propose heuristic techniques that further improve the efficiency of PSP. We evaluate the performance of proposed methods using the federated learning specific FEMNSIT dataset. The evaluation results show that PSP can effectively achieve good balance between system efficiency and model accuracy, mitigating the challenge of heterogeneity in federated learning.

Emerging Machine Learning (ML) techniques, such as Deep Neural Network, are widely used in today's applications and services. However, with social awareness of privacy and personal data rapidly rising, it becomes a pressing and challenging societal issue to both keep personal data private and benefit from the data analytics power of ML techniques at the same time. In this paper, we argue that to avoid those costs, reduce latency in data processing, and minimise the raw data revealed to service providers, many future AI and ML services could be deployed on users' devices at the Internet edge rather than putting everything on the cloud. Moving ML-based data analytics from cloud to edge devices brings a series of challenges. We make three contributions in this paper. First, besides the widely discussed resource limitation on edge devices, we further identify two other challenges that are not yet recognised in existing literature: lack of suitable models for users, and difficulties in deploying services for users. Second, we present preliminary work of the first systematic solution, i.e. Zoo, to fully support the construction, composing, and deployment of ML models on edge and local devices. Third, in the deployment example, ML service are proved to be easy to compose and deploy with Zoo. Evaluation shows its superior performance compared with state-of-art deep learning platforms and Google ML services.

Machine Learning (ML) techniques have begun to dominate data analytics applications and services. Recommendation systems are the driving force of online service providers such as Amazon. Finance analytics has quickly adopted ML to harness large volume of data in such areas as fraud detection and risk-management. Deep Neural Network (DNN) is the technology behind voice-based personal assistance, self-driving cars [1], image processing [3], etc. Many popular data analytics are deployed on cloud computing infrastructures. However, they require aggregating users’ data at central server for processing. This architecture is prone to issues such as increased service response latency, communication cost, single point failure, and data privacy concerns.

Scoped-flooding is used for content discovery in a broad networking context and it has significant impact on the design of caching algorithms in a communication network. Despite its wide usage, a thorough analysis on how scoped-flooding affects a network's performance, e.g., caching and content discovery efficiency, is missing. To develop a better understanding, we first model the behavior of scoped-flooding by the help of a theoretical model on network growth and utility. Next, we investigate the effects of scoped-flooding on various topologies in information-centric networks (ICNs). Using the proposed ring model, we show that flooding can be constrained within a small neighborhood to achieve most of the gains which come from areas with relatively low growth rate, i.e., the network edge. We also study two flooding strategies and compare their behaviors. Given that caching schemes favor more popular items in competition for cache space, popular items are expected to be stored in diverse parts of the network compared to the less popular items. We propose to exploit the resulting divergence in availability along with the routers' topological properties to fine tune the flooding radius. Our results shed light on designing both efficient content discovery mechanism and effective caching algorithms for future ICN.

Recent research has shown a substantial active presence of bots in online social networks (OSNs). In this paper we utilise our previous work (Stweeler) to comparatively analyse the usage and impact of bots and humans on Twitter, one of the largest OSNs in the world. We collect a large-scale Twitter dataset and define various metrics based on tweet metadata. Using a human annotation task we assign 'bot' and 'human' ground-truth labels to the dataset, and compare the annotations against an online bot detection tool for evaluation. We then ask a series of questions to discern important behavioural characteristics of bots and humans using metrics within and among four popularity groups. From the comparative analysis we draw differences and interesting similarities between the two entities, thus paving the way for reliable classification of bots, and studying automated political infiltration and advertisement campaigns.

Information-centric networking (ICN) is a popular research topic. At its heart is the concept of in-network caching. Various algorithms have been proposed for optimizing ICN caching, many of which rely on collaborative principles, i.e. multiple caches interacting to decide what to store. Past work has assumed altruistic nodes that will sacrifice their own performance for the global optimum. We argue that this assumption is insufficient and oversimplifies the reality. We address this problem by modeling the in-network caching problem as a Nash bargaining game. We develop optimal and heuristic caching solutions that consider both performance and fairness. We argue that only algorithms that are fair to all parties involved in caching will encourage engagement and cooperation. Through extensive simulations, we show our heuristic solution, FairCache, ensures that all collaborative caches achieve performance gains without undermining the performance of others.

Efficient index structures for fast approximate nearest neighbor queries are required in many applications such as recommendation systems. In high-dimensional spaces, many conventional methods suffer from excessive usage of memory and slow response times. We propose a method where multiple random projection trees are combined by a novel voting scheme. The key idea is to exploit the redundancy in a large number of candidate sets obtained by independently generated random projections in order to reduce the number of expensive exact distance evaluations. The method is straightforward to implement using sparse projections which leads to a reduced memory footprint and fast index construction. Furthermore, it enables grouping of the required computations into big matrix multiplications, which leads to additional savings due to cache effects and low-level parallelization. We demonstrate by extensive experiments on a wide variety of data sets that the method is faster than existing partitioning tree or hashing based approaches, making it the fastest available technique on high accuracy levels.

Novel network systems need to be carefully evaluated before their actual deployments in developing regions. However, large-scale network experiment is a challenging task. Simulations, emulations, and real-world testbeds all have their advantages and disadvantages. In this paper we present LiteLab, a light-weight platform specialized for large-scale networking experiments. We cover in detail its design, key features, and architecture. We also perform an extensive evaluation of Lite-Lab's performance and accuracy and show that it is able to both simulate network parameters with high accuracy, and also able to scale up to very large networks. LiteLab is flexible, easy to deploy, and allows researchers to perform large-scale network experiments with a short development cycle. We have used LiteLab for many different kinds of network experiments and are planning to make it available for others to use as well.

Caching is a core principle of information-centric networking (ICN). Many novel algorithms have been proposed for enabling ICN caching, many of which rely on collaborative principles, i.e. multiple caches interacting to decide what to store. Past work has assumed entirely altruistic nodes that will sacrifice their own performance for the global optimum. In this paper, we argue that this assumption is flawed. We address this problem by modelling the in-network caching problem as a Nash bargaining game. We develop optimal and heuristic caching solutions that explicitly consider both performance and fairness. We argue that only algorithms that are fair to all parties will encourage engagement and cooperation. Through extensive simulations, we show our heuristic solution, FairCache, ensures that all collaborative caches achieve performance gains without undermining the performance of others.

The ICT industry has come under criticism as being one of the major energy consumers to exacerbate high global carbon emissions. Meanwhile, using renewable energy to power ICT infrastructure is becoming an attractive solution and is gaining its momentum due to the recent breakthroughs of converting solar and wind energies as power sources at competitive costs. Although significant amounts of fossil fuel based-energy can be saved by allowing network devices (e.g., routers and line-cards) to be set to sleep, this optimization approach comes at a price of degrading routing performance, i.e., the quality of service. This paper addresses the problem of minimizing fossil fuel consumption in large Internet Service Provider (ISP) networks, by utilizing a novel gradient-based routing protocol, which favors forwarding packets along routers powered by the highest quantity of renewable energies. Besides favoring renewable energy, the proposed routing protocol can support putting routers to sleep in order to optimize energy consumption while ensuring a minimum degradation in routing performance. Through our evaluation utilizing real meteorological data, our proposed solution has demonstrated a massive reduction of fossil fuel usage by the network (> 70%) while maintaining the routing performance to a similar level when no energy optimization is applied.

Seamless computing and service sharing in community networks (CNs) have gained momentum due to the emerging technology of community network micro-clouds (CNMCs). However, deploying and running services in CNMCs confront enormous challenges to cope with, such as the dynamic nature of micro-clouds, limited capacity of nodes and links, asymmetric quality of wireless links, geographic singularity based deployment model rather than network QoS based, etc. CNMCs have been increasingly used by network-intensive services which exchange significant amounts of data between nodes, therefore their performance heavily relies on the available bandwidth resource in a network. This paper proposes a novel bandwidth-aware service placement algorithm which aims to replace the current random placement adopted by Guifi.net. Our experimental results show that the proposed BASP algorithm consistently outperforms the random placement in Guifi.net by 35% regarding its bandwidth gain. More promisingly, as the number of services increases, the gain tends to increase accordingly.

The Internet is overloading its users with excessive information flows, so that effective content-based filtering becomes crucial in improving user experience and work efficiency. Latent semantic analysis has long been demonstrated as a promising information retrieval technique to search for relevant articles from large text corpora. We build Kvasir, a semantic recommendation system, on top of latent semantic analysis and other state-of-the-art technologies to seamlessly integrate an automated and proactive content provision service into web browsing. We utilize the processing power of Apache Spark to scale up Kvasir into a practical Internet service. In addition, we improve the classic randomized partition tree to support efficient indexing and searching of millions of documents. Herein we present the architectural design of Kvasir, the core algorithms, along with our solutions to the technical challenges in the actual system implementation.

Content-Centric Network (CCN) shifts the Internet from point-to-point paradigm to receiver-driven data-centric paradigm. While it tries to solve many problems in the current Internet and opens the door to many novel applications, it also leaves many challenges unanswered, e.g., mobility support and mobile content publishing and dissemination. In this paper, we show how a greedy routing can be implemented in CCN architecture to support mobility. This allows for efficient content publisher mobility and supports seamless handoffs for interactive connections. We present our solution - MobiCCN, and evaluate it thoroughly in realistic network topologies to show it outperforms other popular mobility schemes.

Peer-to-peer networks have been quite thoroughly measured over the past years, however it is interesting to note that the BitTorrent Mainline DHT has received very little attention even though it is by far the largest of currently active overlay systems, as our results show. As Mainline DHT differs from other systems, existing measurement methodologies are not appropriate for studying it. In this paper we present an efficient methodology for estimating the number of active users in the network. We have identified an omission in previous methodologies used to measure the size of the network and our methodology corrects this. Our method is based on modeling crawling inaccuracies as a Bernoulli process. It guarantees a very accurate estimation and is able to provide the estimate in about 5 seconds. Through experiments in controlled situations, we demonstrate the accuracy of our method and show the causes of the inaccuracies in previous work, by reproducing the incorrect results. Besides accurate network size estimates, our methodology can be used to detect network anomalies, in particular Sybil attacks in the network. We also report on the results from our measurements which have been going on for almost 2.5 years and are the first long-term study of Mainline DHT.

Caching is a key component of information-centric networking, but most of the work in the area focuses on simple en-route caching with limited cooperation between the caches. In this paper we model cache cooperation under a game theoretical framework and show how cache cooperation policy can allow the system to converge to a Pareto optimal configuration. Our work shows how cooperation impacts network caching performance and how it takes advantage of the structural properties of the underlying network.

Distributed hash tables (DHT) are a key building block for modern P2P content-distribution system, for example in implementing the distributed tracker of BitTorrent Mainline DHT. DHTs, due to their fully distributed nature, are known to be vulnerable to certain kinds of attacks and different kinds of defenses have been proposed against these attacks. In this paper, we consider two kinds of attacks on a DHT, one already known attack and one new kind of an attack, and show how they can be targeted against Mainline DHT. We complement them by an extensive measurement study using honeypots which shows that both attacks have been going on for a long time in the network and are still happening. We present numbers showing that the number of sybils in the Mainline DHT network is increasing and is currently around 300,000. We analyze the potential threats from these attacks and propose simple countermeasures against them.

In-network caching of content is a popular technique for eliminating redundant traffic from the network and improve the performance of network applications. In this paper we present a novel cooperative caching strategy to improve performance of in-network caches. Our cooperative scheme is composed of an admission policy for the incoming data and a content exchange protocol between neighbor network caches to improve the search zone. The admission policy enforces that a previously cached data is not unnecessary replicated in other caches, resulting in more space for new data. The content exchange protocol allows for exchange on cached data, increasing the hit rate for incoming requests. The benefits are twofold: first, we reduce the redundant content caching in the network, and second, we improve the hit rate by informing the content cached in the nearby caches. As a proof-of-concept, we have implemented a prototype and evaluated its performance using different large-scale topologies against standard non-cooperative caching algorithms. Our numerical results show that both admission and content exchange policies yield large performance gains over standard algorithms.