Effective Usage of Computational Trust Models in Rational Environments

Computational reputation-based trust models using statistical learning have been intensively studied for distributed systems where peers behave maliciously. However practical applications of such models in environments with both malicious and rational behaviors are still very little understood. In this paper, we study the relation between their accuracy measures and their ability to enforce cooperation among participants and discourage selfish behaviors. We provide theoretical results that show the conditions under which cooperation emerges when using computational trust models with a given accuracy and how cooperation can be still sustained while reducing the cost and accuracy of those models. Specifically, we propose a peer selection protocol that uses a computational trust model as a dishonesty detector to filter out unfair ratings. We prove that such a model with reasonable misclassification error bound in identifying malicious ratings can effectively build trust and cooperation in the system, considering rationality of participants. These results reveal two interesting observations. First, the key to the success of a reputation system in a rational environment is not a sophisticated trust learning mechanism, but an effective identity management scheme to prevent whitewashing behaviors. Second, given an appropriate identity management mechanism, a reputation-based trust model with a moderate accuracy bound can be used to enforce cooperation effectively in systems with both rational and malicious participants. As a result, in heterogeneous environments where peers use different algorithms to detect misbehavior of potential partners, cooperation may still emerge. We verify and extend these theoretical results to a variety of settings involving honest, malicious and strategic players through extensive simulation. These results will enable a much more targeted, cost-effective and realistic design for decentralized trust management systems, such as needed for peer-to-peer, electronic commerce or community systems.

High Quality P2P Service Provisioning via Decentralized Trust Management

Hung Le Vu

Trust management is essential to fostering cooperation and high quality service provisioning in several peer-to-peer (P2P) applications. Among those applications are customer-to-customer (C2C) trading sites and markets of services implemented on top of centralized infrastructures, P2P systems, or online social networks. Under these application contexts, existing work does not adequately address the heterogeneity of the problem settings in practice. This heterogeneity includes the different approaches employed by the participants to evaluate trustworthiness of their partners, the diversity in contextual factors that influence service provisioning quality, as well as the variety of possible behavioral patterns of the participants. This thesis presents the design and usage of appropriate computational trust models to enforce cooperation and ensure high quality P2P service provisioning, considering the above heterogeneity issues. In this thesis, first I will propose a graphical probabilistic framework for peers to model and evaluate trustworthiness of the others in a highly heterogeneous setting. The framework targets many important issues in trust research literature: the multi-dimensionality of trust, the reliability of different rating sources, and the personalized modeling and computation of trust in a participant based on the quality of services it provides. Next, an analysis on the effective usage of computational trust models in environments where participants exhibit various behaviors, e.g., honest, rational, and malicious, will be presented. I provide theoretical results showing the conditions under which cooperation emerges when using trust learning models with a given detecting accuracy and how cooperation can still be sustained while reducing the cost and accuracy of those models. As another contribution, I also design and implement a general prototyping and simulation framework for reputation-based trust systems. The developed simulator can be used for many purposes, such as to discover new trust-related phenomena or to evaluate performance of a trust learning algorithm in complex settings. Two potential applications of computational trust models are then discussed: (1) the selection and ranking of (Web) services based on quality ratings from reputable users, and (2) the use of a trust model to choose reliable delegates in a key recovery scenario in a distributed online social network. Finally, I will identify a number of various issues in building next-generation, open reputation-based trust management systems as well as propose several future research directions starting from the work in this thesis.

EPFL2010

Live Streaming with Gossip

Maxime Monod

Peer-to-peer (P2P) architectures have emerged as a popular paradigm to support the dynamic and scalable nature of distributed systems. This is particularly relevant today, given the tremendous increase in the intensity of information exchanged over the Internet. A P2P system is typically composed of participants that are willing to contribute resources, such as memory or bandwidth, in the execution of a collaborative task providing a benefit to all participants. File sharing is probably the most widely used collaborative task, where each participant wants to receive an individual copy of some file. Users collaborate by sending fragments of the file they have already downloaded to other participants. Sharing files containing multimedia content, files that typically reach the hundreds of megabytes to gigabytes, introduces a number of challenges. Given typical bandwidths of participants of hundreds of kilobits per second to a couple of megabits per second, it is unacceptable to wait until completion of the download before actually being able to use the file as the download represents a non negligible time. From the point of view of the participant, getting the (entire) file as fast as possible is typically not good enough. As one example, Video on Demand (VoD) is a scenario where a participant would like to start previewing the multimedia content (the stream), offered by a source, even though only a fraction of it has been received, and then continue the viewing while the rest of the content is being received. Following the same line of reasoning, new applications have emerged that rely on live streaming: the source does not own a file that it wants to share with others, but shares content as soon as it is produced. In other words, the content to distribute is live, not pre-recorded and stored. Typical examples include the broadcasting of live sports events, conferences or interviews. The gossip paradigm is a type of data dissemination that relies on random communication between participants in a P2P system, sharing similarities with the epidemic dissemination of diseases. An epidemic starts to spread when the source randomly chooses a set of communication partners, of size fanout, and infects them, i.e., it shares a rumor with them. This set of participants, in turn, randomly picks fanout communication partners each and infects them, i.e., share with them the same rumor. This paradigm has many advantages including fast propagation of rumors, a probabilistic guarantee that each rumor reaches all participants, high resilience to churn (i.e., participants that join and leave) and high scalability. Gossip therefore constitutes a candidate of choice for live streaming in large-scale systems. These advantages, however, come at a price. While disseminating data, gossip creates many duplicates of the same rumor and participants usually receive multiple copies of the same rumor. While this is obviously a feature when it comes to guaranteeing good dissemination of the rumor when churn is high, it is a clear disadvantage when spreading large amounts of multimedia data (i.e., ordered and time-critical) to participants with limited resources, namely upload bandwidth in the case of high-bandwidth content dissemination. This thesis therefore investigates if and how the gossip paradigm can be used as a highly effcient communication system for live streaming under the following specific scenarios: (i) where participants can only contribute limited resources, (ii) when these limited resources are heterogeneously distributed among nodes, and (iii) where only a fraction of participants are contributing their fair share of work while others are freeriding. To meet these challenges, this thesis proposes (i) gossip++: a gossip-based protocol especially tailored for live streaming that separates the dissemination of metadata, i.e., the location of the data, and the dissemination of the data itself. By first spreading the location of the content to interested participants, the protocol avoids wasted bandwidth in sending and receiving duplicates of the payload, (ii) HEAP: a fanout adaptation mechanism that enables gossip to adapt participants' contribution with respect to their resources while still preserving its reliability, and (iii) LiFT: a protocol to secure high-bandwidth gossip-based dissemination protocols against freeriders.

EPFL2010

Structure-preserving approaches and data-driven closure modeling for model order reduction

Nicolò Ripamonti

In this thesis, we propose model order reduction techniques for high-dimensional PDEs that preserve structures of the original problems and develop a closure modeling framework leveraging the Mori-Zwanzig formalism and recurrent neural networks. Since high-fidelity approximations of PDEs often result in a large number of degrees of freedom, the need for iterative evaluations for numerical optimizations and rapid feedback is computationally challenging.The first part of this thesis is devoted to conserving the high-dimensional equation's invariants, symmetries, and structures during the reduction process. Traditional reduction techniques are not guaranteed to yield stable reduced systems, even if the target problem is stable. In the context of fluid flows, the skew-symmetric structure of the problem entails the preservation of the kinetic energy of the system. By preserving the same structure at the level of the reduced model, we obtain enhanced stability, and accuracy and the reduced model acquires physical significance by preserving a surrogate of the energy of the original problem. Next, we focus on Hamiltonian systems, which, being driven by symmetry, are a source of great interest in the reduction community. It is well known that the breaking of these symmetries in the reduced model is accompanied by a blowup of the system energy and flow volume. In this thesis, geometric reduced models for Hamiltonian systems are further developed and combined with the dynamically orthogonal methods, addressing the poor reducibility in time of advection-dominated problems. The reduced solution is expressed as a linear combination of a finite number of modes and coincides with the symplectic projection of the high-fidelity Hamiltonian problem onto the tangent space of the approximating manifold. An error surrogate is used to monitor the approximation ability of the reduced model and make a change in the rank of the approximating system if necessary. The method is further developed through a combination of DEIM and DMD to reduce non-polynomial nonlinearities while preserving the symplectic structure of the problem and applied to the Vlasov-Poisson system.In the second part of the thesis, we consider several data-driven methods to address the poor accuracy in the under-resolved regime for Galerkin reduced models via a closure term. The closure term is developed systematically from the Mori-Zwanzig formalism by introducing projection operators on the spaces of resolved and unresolved scales, thus resulting in an additional memory integral term. The interaction between different scales turns out to be nonlocal in time and dominated by a high-dimensional orthogonal dynamics equation, which cannot be solved precisely and efficiently. Several classical methods in the field of statistical mechanics are used to approximate the memory term, exploiting the finiteness of the memory kernel support. We conclude this thesis by showing through numerical experiments how long short-term memory networks, i.e., machine learning structures characterized by feedback connections, represent a valid tool for approximating the additional memory term.

EPFL2022

High Quality P2P Service Provisioning via Decentralized Trust Management

Hung Le Vu

EPFL2010

Structure-preserving approaches and data-driven closure modeling for model order reduction

Nicolò Ripamonti

EPFL2022

Live Streaming with Gossip

Maxime Monod

EPFL2010