Dissertation Defense: Md Shoaib Ahmed

Title: TRUSTWORTHY AI FOR INFORMATION ECOSYSTEMS: ADVERSARIAL ATTACK, TRUST-AWARE DESIGN, AND ROBUSTNESS EVALUATION

Program: Computing PhD

Committee Chair: Francesca Spezzano

Committee: Francesca Spezzano, Jun Zhuang, Nasir Eisty

Abstract: The rapid spread of misinformation in digital spaces threatens public trust and the reliability of online information, while the growing use of AI-driven systems in social media has introduced significant challenges to the trustworthiness of digital information ecosystems. Social media platforms increasingly rely on AI-based fake news detectors and recommendation algorithms to manage the information users encounter. However, these systems remain vulnerable to adversarial manipulation, and their reliability under different attack conditions is not yet fully understood.

This dissertation investigates trustworthiness in AI-driven information ecosystems across three interconnected dimensions: exposing adversarial vulnerabilities in misinformation detection, developing trust-aware recommendation algorithms, and evaluating the robustness of news recommender systems against large language model (LLM) driven textual attacks.

First, we introduce an Explainable AI (XAI)-guided adversarial attack framework that combines SHAP-based token attribution with LL

Ms to generate contextually relevant adversarial comments targeting fake news detection systems. Experimental findings reveal that explainability techniques intended to improve model transparency can also be exploited to manipulate classification outcomes, highlighting the dual-use risks of XAI and the varying robustness of detection architectures.

Second, we address misinformation at the distribution layer by introducing MA-Graph

Rec, a misinformation-aware recommendation architecture that incorporates trustworthy social network relationships and user-neighbor information into graph-based recommendation modeling. Experimental evaluations show that integrating trustworthy user information reduces misinformation exposure at the recommendation stage and limits its subsequent dissemination through social networks, while accounting for personalized recommendation quality.

Finally, we investigate the adversarial robustness of neural news recommender systems through LEXA (LE

Xicon-guided Attack), a framework that combines interpretable linguistic features, explainable machine learning, and LLM-based news rewriting to manipulate recommendation outcomes. By identifying linguistic patterns associated with favorable rankings and using them to guide meaning-preserving modifications of news titles and abstracts, LEXA demonstrates that targeted textual changes can improve recommendation rankings under certain conditions. However, the findings also reveal that ranking improvements do not consistently translate into increased recommendation visibility and that attack effectiveness varies across recommender architectures.

In summary, this dissertation contributes to the develo