An MSc dissertation examining how reliably transformer-based authorship verification (AV) models detect the same author across text domains (news vs. tweets) and under LLM-based adversarial rewriting. Using the CROSSNEWS dataset, it quantifies how vulnerable current AV systems are to style obfuscation and impersonation attacks — with implications for AI safety and misuse detection.

EEG signals are inherently challenging to interpret, even for trained clinicians. This study investigates whether deep learning models can accurately detect harmful brain activity from EEG data, while also addressing the interpretability of model predictions using Grad-CAM.

A browser-based tool that detects AI voice-cloning (deepfake) phone scams, built on Gemma 4's multimodal audio encoder. Winner of the Safety & Trust track at the Gemma 4 Good Hackathon (Google DeepMind × Kaggle). Repurposes a frozen audio encoder as an anti-spoofing feature extractor, reaching 5.20% EER with only a 1.5MB classification head.