4pm, 1.02 MPEB

We will introduce the concept of "On Demand Connectivity Sharing," which we build on top of User-Provided Networks (UPNs). UPNs were recently proposed as a new connectivity paradigm, according to which home-users share their broadband Internet connection with roaming guests. We enhance this paradigm with incentives, rules and policies, based on which: (i) home-users provide on-demand connectivity only (i.e., they do not explicitly allocate a portion of their bandwidth) and (ii) guest-users utilize resources that remain unexploited from the respective home-users.

Our findings indicate the following: (i) ubiquitous Internet connectivity (in densely populated, urban areas) is already possible, since the necessary infrastructure is in place, (ii) resource sharing can take place in a seamless manner for the home-user and (iii) mobile-, guest-users can have acceptable performance for short data-transfers, such as e-mailing, surfing, map-download or photo uploads.

2pm, 6.12 MPEB

This paper proposes a novel and efficient cooperative leader election algorithm for master/slave mobile ad hoc networks. The algorithm relies on collecting and re-distributing information amongst local nodes in order to find the leader. It is based on the assumption that if this process is repeated sufficiently then the algorithm will converge towards a unique leader. It is shown that the proposed mechanism outperforms existing algorithms in terms of time complexity and response to node mobility. The algorithm was simulated for Bluetooth ad hoc networks, which, by default, rely on a master/slave architecture, however, the cooperative approach could be adapted to any network that exhibits the master/slave configuration such as clustered ad hoc networks or ZigBee-based sensor networks.

6.12 MPEB, 2pm

The assessment and use of reputations to make decisions that may involve risk - in the form of a loss in utility - raises a number fundamental questions regarding the properties of reputation models. Amongst these is what types of equilibrium exist to support a cooperative outcome. In this talk we show how the characteristics of the interaction model shape the type of feasible equilibrium and highlight the short comings this raises for making decisions based on reputations.

2pm, 6.12 MPEB

The rise of Web 2.0 has emphasized new issues and challenges for the research community. The amount of data produced and consumed by users has been exponentially increasing under the incentives exerted by the proliferation of information sharing communities. An enormous amount of pictures, blog entries, messages and media resources are flowing and are being shared all over the network, backed also by the growing connectivity of new generation mobile devices.

The new pervasive technologies allow everyone to access and load their contents and resources online, regardless of their location. Disseminating, disclosing and, in general, sharing data have become mainstream activities.
In this scenario, a new trend called social (or folksonomic) tagging has emerged as the most popular way to describe, categorize, search and navigate information. Folksonomies allow users to associate data with customized tags, without forcing information to be categorized in pre-existent hierarchies: users uploading resources on the web are free to label data with the unlimited vocabulary of spoken language.

However, as tags are informally defined, continually changing and ungoverned, social tagging has often been criticized for lowering, rather than increasing, the efficiency of searching. New ways of retrieving information matching users’ requests must be investigated.

Our work aims at facilitating the problem of information retrieval for Web 2.0 scenarios in which items are not categorized into hierarchies, but defined with descriptive tags, and explicit ratings are not available to understand users’ preferences: interests are implicitly inferred from bookmarked items.
By means of an extensive set of experiments using a variety of large-scale real datasets (CiteULike, MovieLens, Bibsonomy, Delicious), the proposed solution will be evaluated in terms of accuracy (how much returned data accomplish users’ interests), completeness (how much relevant data is retrieved) and scalability (how the solution can handle growing amount of content).

2pm, 6.12 MPEB

Thanks to advances in the computing capabilities and added functionalities of modern mobile devices, creating digital media on the move has never been so easy and popular. However, the distribution of such content still represents a challenge: the coverage/stability of 3G networks is limited, and their cost high. As a result, distributing content using Delay Tolerant Networks (DTNs) has become an attractive alternative.

Most of the DTN routing protocols proposed in the literature have been exploiting users' mobility patterns, in order to maximise the delivery probability, while minimising the overall network overhead (e.g., number of replicas in the system, messages' path length). Common to all these protocols has been the assumption that devices are willing to participate in the content distribution network; however, because of battery constraints, participation cannot be taken for granted, especially if the very same subset of devices are continuously selected as content carriers, simply because of their mobility properties. Indeed, we demonstrate that state-of-the-art DTN routing protocols distribute load in a highly unfair manner, with detrimental effects on delivery once the assumption of unconditional participation is lifted.

To overcome this problem, we propose CoHabit, a fair content dissemination protocol for participatory DTNs.  By adopting a source-based approach to routing, local estimations of the network workload can be computed and used to select routes that favour the least loaded portion of the network, thus preventing unduly load distribution. We demonstrate, by means of simulation using real large-scale mobility traces, that CoHabit achieves high delivery without compromising fairness.

2pm, 6.12 MPEB

The talk considers the problem of synthesising accurate dynamic models of a miniature rotorcraft based on minimal knowledge about the flying machine. The approach is based on the idea of building models that predict the dynamic accelerations affecting the platform, and is implemented using evolutionary programming. A co-evolutionary framework is shown to be very effective in identifying both the structure and the parameters of the nonlinear models from real flight data. The modelling method is demonstrated on a miniature quadrotor rotorcraft; the automatically obtained models are used to develop a controller capable of autonomous flight.

2pm, 6.12 MPEB

2pm, 6.12 MPEB (Note: To Be Rescheduled)

In wireless sensor networks (WSN), physical topology control is a strategy used to construct a reliable network while minimising the transmission power used by each node. In particular, we would like to maintain network connectivity and provide some redundancy to the network, while minimising the total energy usage and average node degree to increase network capacity in the network. We present Subgraph Topology Control (STC), a simple, distributed iterative algorithm, which constructs networks using Subgraph Number parameter obtained by exchanging neighbour tables between immediate nodes. Subgraph Number provides a good insight into the local connectivity using only node IDs. We show that networks constructed by STC have lower node degree and achieve better connectivity compare to k-neighbours, which also uses node degree for topology control. The iterative nature means that although STC initially takes longer to reach stability, only nodes that are in vicinity to a topology change are involved in an update. Moreover, we demonstrate STC is self-healing and can adapt to dynamic topological changes.

2pm, 6.12 MPEB

Recommender systems are vulnerable to attack: malicious users may deploy a set of sybils to inject ratings in order to damage or modify the output of Collaborative Filtering (CF) algorithms. Previous work focuses on designing sybil profile classification algorithms, which operate independently of CF, and aim to find the current sybils each time they are run. These methods, however, assume that the full sybil profiles have already been input to the system. Deployed recommender systems, on the other hand, operate over time: recommendations may be damaged as sybils inject profiles (rather than only when all the malicious ratings have been input), and system administrators may not know when their system is under attack. In this work, we formalise the notion of a temporal sybil attack, and propose and evaluate methods for monitoring global, user and item behaviour over time in order to detect rating anomalies that reflect an ongoing attack. We conclude by discussing the consequences of our temporal defenses, and how attackers may design ramp-up attacks in order to circumvent them.

Slides are now available on slideshare.

2pm, MPEB 6.12

I will give an overview of topics I worked on and I will give a high level overview of different machine learning techniques. In particular, I will talk about classfication/ regression methods, about methods for time series processing, about reinforcement Learning and about Bandit algorithms. The hope is that people get a feeling of what problems can be tackled with machine learning techniques and what techniques might be worth a try.