Identificazione e Rilevamento Dell'Attività Dei Troll Nei Social Network Online
In sintesi
I social network online (OSN) hanno trasformato il modo in cui si accede alle notizie, si condividono opinioni, si fanno affari e politica. L'accuratezza, la veridicità e l'autenticità dei dati condivisi sono ingredienti necessari per un sano ecosistema. Tuttavia, recentemente, l’enorme crescita di account maligni, che condividono informazioni false e provocatorie per influenzare l'opinione pubblica, ha creato conflitti su varie questioni, con enormi effetti negativi. Ad oggi, i tentativi di rilevare tali account, bot (gestiti da SW) e troll (umani provocatori), sono stati poco efficaci.
Soggetto e obiettivo
Il progetto propone una soluzione computazionale innovativa per il rilevamento automatico di troll negli OSN che sia agnostico sia in termini di piattaforma, che per l'origine e lo scopo degli utenti malintenzionati. La soluzione ha l’obiettivo di essere generalizzabile a diversi OSN e scenari (come: nel contesto politico, sanitario o sociale). Questo grazie ad una metodologia originale, che non si limita ad identificare i troll, ma mira anche a capire e caratterizzare la loro attività nefasta. La logica è quella di dedurre gli incentivi alla base del comportamento dei troll e utilizzare tali elementi per distinguerli dagli altri account.
Contesto socio-scientifico
L'impatto previsto è enorme per l’integrità delle attività su OSN e, quindi per la salvaguardia dei capisaldi della nostra società, dalla democrazia alla salute pubblica. Infatti, questo sistema di rilevamento potrà essere utilizzato, anche dai gestori di OSN, in contesti diversi come la manipolazione dell'opinione pubblica, la disinformazione, le molestie online e il cyber-bullismo, la manomissione di dibattiti, che possono creare panico o negazionismo in casi di emergenza (come per il COVID-19).
Online Social Networks (OSNs) are computer-based technologies that enable users to create content and entertain social relationships. OSNs have rapidly grown from an aggregation channel to a global phenomenon that has elicited a paradigm shift of our society, by transforming the way people access the news, share opinions, make business and politics. In such a scenario, accuracy, truthfulness, and authenticity of the shared content are necessary ingredients to maintain a healthy online discussion. However, in recent times, OSNs have been dealing with a considerable growth of malicious and fake accounts, which purposely undermine the integrity of online conversations by sharing false and inflammatory information in OSN platforms to influence the public opinion and sow conflict on social or political issues. Recent studies found evidence that these malicious actors play a disproportionate role in manipulation and misinformation campaigns globally, highlighting the pitfalls of OSNs and their potential harms to several constituents of our society, ranging from politics to public health. Bots (i.e., software-controlled accounts) and trolls (i.e, human operators often tied with political movements) represent the most recognized categories of malicious entities acting in such efforts. Although the attempt of OSN providers to purge their platforms, the nefarious activity of bots and trolls has not entirely stopped. While the detection of bots has established solutions, the automated identification of troll accounts represents an open challenge for the research community. Existing approaches analyze the activity of accounts tied to Russia’s troll farm and involved in the Twitter online discussion during the 2016 US Presidential election. To enable the detection of such trolls, these efforts use language cues and Twitter accounts’ metadata. Such solutions, other than being limited to a unique OSN, exploits linguistic markers strictly related to the trolls under analysis.In this project, we aim to launch an innovative computational tool for automatically detecting troll accounts in OSNs that is agnostic both on the online platform under scrutiny and on the origin and purpose of the malicious users. The targeted solution strives to be generalizable to different OSNs and scenarios (e.g., in the political, health, or social context). The novelty of our solution is also in the proposed methodology, which does not merely target to identify troll accounts in online platforms, but also aim to understand and characterize their malicious activity. The rationale is to infer a set of online incentives that may steer trolls’ behavior and utilize such cues for distinguishing them from non-troll accounts. To perform such analysis, we intend to rely on Inverse Reinforcement Learning (IRL), which is a machine learning framework that has the main goal of finding the motivation (in the form of rewards) behind an observed behavior. In our scenario, we plan to employ IRL to infer the rewards that could have led troll and non-troll users to perform their online activity. We then consider to exploit the estimated rewards as features of a supervised learning algorithm aimed at classifying such accounts. The differences in the predictive features between the two classes of accounts might enable a principled understanding of the distinctive behaviors reflecting the incentives trolls and non-trolls respond to. While these objectives represent ambitious and critical challenges in the race towards healthy online ecosystems, our pioneering approach, given its generalizable nature, can potentially impact and curb the activity of malicious actors along different dimensions, not only related to the manipulation of public opinion in the political context, but also, and more generally, in the fight against every form of online harassment and abuse in OSNs (e.g., cyberbullying).
Last updated:11.06.2022
