Trust and reputation systems: detection of malicious agents and a novel equilibrium problem

This thesis is devoted to the analysis and development of trust and reputation systems, focusing on the study of accurate models for the computation of reputation, especially in virtual contexts such as social networks. In these environments, where users share feedback based on their experiences, we investigated the dynamics that influence reputation formation. By analysing what causes users to assign trust and how reputation is perceived, we have proposed advanced models that aim at accurately capturing the complexity of these interactions. In the first part of the thesis we focused on model development with an engineering approach, adopting a data processing engineering perspective. After a brief introduction to the state of the art and the main definitions, such as trust, reputation, multi-agents system and agents organization, we delved into the field by presenting two key algorithms. These algorithms were specifically designed to identify malicious colluding users present in the virtual environment. Through these tools, we were able to more precisely identify users who act maliciously and who could compromise the integrity of virtual environments. We have examined the well known Eigentrust algorithm. This algorithm is recognized as one of the most effective solution to measure the reputation in a set of social agents. We have highlighted a possible limitation in the Eingentrust algorithm, observing that it increases the reputation of all the pre-trusted agents, regardless of their reliability. This way, the method is capable to effectively recognize colluding agents, but producing the side effect to flattening the differences, in terms of reliability, between honest agents. To address the problem above, as first algorithm, we have proposed a different strategy, based on a decrement of the fraudulent trust values that colluding agents mutually exchange. Our strategy introduces the advantage, with respect to Eigentrust, of estimating the reputation values of the honest actors in a manner more adherent to the actual reliability of these agents. Instead, the second proposed algorithm was designed for the detection of colluding agents divided into multiple clusters. It combines the EigenTrust algorithm with a clustering procedure, grouping agents based on their reputation scores. A crucial result of this first part was the ability to calculate the actual reputation of honest users, taking into account the fraudulent actions and interactions present in such contexts. In the second part of the thesis, we dedicated to study trust and reputation models by adopting a mathematical approach, employing the theory of variational inequalities as an analysis tool. This methodology was applied in a virtual environment, in which users evaluated specific objects. Our analysis focused on how these mathematical models could be e_ectively applied to understand and improve the concept of reputation in digital environments. By formulating trust and reputation as variational problems, this approach gives us a new perspective on understanding mechanisms that govern the creation of trust. We demonstrate that the equilibrium conditions can be formulated as a variational inequality problem and we provide a novel alternative formulation. We explored the robustness of our model across different conditions, introducing significant variations in both user trustworthiness and initial reputation of objects. This sensitivity analysis allowed us to evaluate how the system responded to changes in key variables, providing valuable insights into its adaptability and reliability in diverse contexts.

In questa tesi ci siamo occupati dell'analisi e dello sviluppo di sistemi di Trust e Reputation, concentrandosi sullo studio di modelli accurati per il calcolo della reputazione, soprattutto in contesti virtuali come i social network. Nella prima parte della tesi, ci siamo concentrati sullo sviluppo di modelli con un approccio ingegneristico, attraverso l'analisi e l'elaborazione dei dati in nostro possesso. Abbiamo esaminato il noto algoritmo Eigentrust, riconosciuto come una delle soluzioni più efficaci per misurare la reputazione in un insieme di agenti sociali. Abbiamo evidenziato però un possibile limite dell'algoritmo Eingentrust, osservando che esso aumenta la reputazione di tutti gli agenti pre-trusted, indipendentemente dalla loro affidabilità. Per affrontare questo problema, come primo algoritmo, abbiamo proposto una strategia basata su una diminuzione dei valori di fiducia fraudolenti che gli agenti collusi si scambiano reciprocamente. Il secondo algoritmo proposto, invece, è stato progettato per il rilevamento di agenti collusi suddivisi in più cluster. Combina l'algoritmo EigenTrust con una procedura di clustering, raggruppando gli agenti in base ai loro punteggi di reputazione. Nella seconda parte della tesi invece, ci siamo dedicati allo studio dei modelli di Trust e Reputation adottando un approccio matematico, utilizzando la teoria delle disequazioni variazionali come strumento di analisi. Questa metodologia è stata applicata in un ambiente virtuale, in cui gli utenti hanno valutato oggetti specifici. La nostra analisi si è concentrata su come questi modelli matematici potrebbero essere applicati efficacemente per comprendere e migliorare il concetto di reputazione negli ambienti digitali. Abbiamo esplorato l'efficienza del nostro modello in diverse condizioni, introducendo variazioni significative sia nell'affidabilità dell'utente che nella reputazione iniziale degli oggetti. Questa analisi di sensibilità ci ha permesso di valutare come il sistema ha risposto ai cambiamenti nelle variabili chiave, fornendo informazioni sulla sua adattabilità in diversi contesti.