ProfileShield: Preventing Psychological Profiling and Behavioural Manipulation in Online Social Media

Core PhD Question

How can we prevent AI systems, advertisers, platforms, and malicious actors from constructing psychological profiles of social media users while still allowing meaningful online interaction, personalization, and content discovery?

This PhD investigates how online social media data can be used to infer a person’s personality, emotions, beliefs, vulnerabilities, interests, habits, and persuasion triggers. The project focuses on designing AI-based privacy and security mechanisms that can detect, measure, and reduce psychological profiling risks before they are exploited for manipulation, microtargeting, discrimination, or behavioral influence.

What We Are Not Planning to Do

This project is not about deleting social media, blocking all personalization, or arguing that online platforms should stop recommending content completely.

It is also not about building another generic fake-news detector, hate-speech classifier, or misinformation detection model.

The project is not focused only on traditional privacy protection such as hiding names, emails, phone numbers, or locations. Psychological profiling can happen even when obvious identifiers are removed.

This project is also not about preventing users from expressing themselves online. The goal is to protect people from invisible inference, manipulation, and exploitation while preserving their ability to communicate naturally.

It is also not about building tools for psychological profiling. Any profiling models used in the project will be used only for defensive measurement, risk evaluation, and privacy protection.

What We Are Planning to Do

This project will study how AI systems can infer psychological and behavioural traits from social media content, including posts, comments, likes, shared links, interaction patterns, writing style, visual content, and engagement behaviour.

The PhD will investigate how large language models, recommender systems, advertising systems, and data brokers may infer sensitive traits such as personality, emotional state, political inclination, social vulnerability, addictive tendencies, financial stress, health concerns, or susceptibility to persuasion.

The project will then design a defensive framework that can measure the psychological profiling risk of a user’s online footprint. For example, the system may analyse whether a set of posts reveals too much about a user’s personality, emotional instability, fears, ideological leaning, or manipulation triggers.

A major part of the project can be the development of a psychological privacy firewall for social media. This firewall can warn users, platforms, or regulators when online content enables high-risk profiling. It may suggest safer rewording, reduce behavioural leakage, perturb sensitive signals, or detect profiling attempts.

The project may also investigate adversarial privacy techniques that allow users to preserve meaning while reducing profileability. For example, a post should still sound natural and useful, but should reveal less about hidden psychological traits.

The project can evaluate these ideas using public social media datasets, simulated user profiles, synthetic behavioural traces, LLM-based profiling attacks, and privacy-risk metrics.

Possible PhD Contribution Expectations

1. Psychological Profiling Threat Model
   A new threat model explaining how AI systems infer personality, emotional state, beliefs, vulnerabilities, and persuasion triggers from social media activity.
2. Profileability Risk Metric
   A measurable privacy-risk score that estimates how easily a user can be psychologically profiled from posts, comments, likes, or behavioural patterns.
3. LLM-Based Profiling Attack Benchmark
   A benchmark that tests how modern AI models can infer psychological traits from social media content, even when direct identifiers are removed.
4. Psychological Privacy Firewall
   A defence system that detects high-risk content, warns users, and suggests privacy-preserving alternatives before posting.
5. Meaning-Preserving Obfuscation Methods
   New methods that reduce psychological leakage while preserving the original message, tone, and communicative value of the user’s content.
6. Anti-Microtargeting Defence Layer
   A mechanism to detect or reduce the use of inferred psychological traits for manipulative advertising, political persuasion, or discriminatory targeting.
7. Explainable Privacy Feedback
   A user-friendly explanation layer that tells users what kind of psychological signals may be exposed, such as stress, fear, impulsiveness, loneliness, ideology, or susceptibility to influence.
8. Evaluation Framework
   A complete evaluation setup measuring privacy protection, text utility, naturalness, user acceptability, profiling resistance, and robustness against adaptive AI profilers.