Boston University
Data-driven approaches for improving the identification of misleading content online
Abstract
dc:description.abstractMisleading content online appears in many forms, spanning false claims that spread rapidly on social networks to craftily designed e-commerce websites defrauding users of money and trust.This thesis aims to build data-driven systems that can substantially improve the automated identification of misleading content online while supporting human-in-the-loop content moderation systems and downstream security systems. I achieve this goal by developing and evaluating four complementary systems that together strengthen platform soft-moderation practices and enable proactive discovery of scam websites on the broader Web. First, I introduce a claim-comprehensive soft-moderation pipeline that uses learning-to-rank and information retrieval techniques to identify posts discussing misleading claims, increasing coverage and consistency of warning labels.Second, I propose an unsupervised, context-aware stance detection framework to distinguish the propagation of a falsehood from its critique or correction, reducing contextual false positives of warning labels. Third, I extend soft moderation beyond text with an efficient reverse image retrieval system that finds visually similar instances of misleading images at scale, enabling multi-modal moderation. Finally, I present a data-driven query-mining and scoring system that allows systematic issuing of search engine queries with a higher likelihood of returning scam websites, accelerating existing security pipelines to discover scam websites earlier, and improving the resource efficiency of downstream detection systems. Across large-scale, heterogeneous datasets capturing real-world events on social media and diverse search engine results on the web, these systems (i)~expand the coverage, context, and accuracy of soft moderation of misleading content online and (ii)~improve the timeliness and yield of discovering online scam websites promoting misleading content.Collectively, these contributions develop a practical toolbox of systems for human-in-the-loop moderation and security systems, demonstrating that targeted, claim-centric, context-aware, and multi-modal pipelines can help make information ecosystems more trustworthy and safe.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Electrical & Computer Engineering
- Grantor
- Boston University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Paudel, Pujan
- Advisor dc:contributor.advisor
-
- Stringhini, Gianluca
- ORCID dc:identifier.orcid
Subjects
dc:subject × 7Rights
dc:rights- Statement dc:rights
-
- Attribution 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/52160