Projects

Can state-of-the-art audio deepfake detectors generalise beyond English? A benchmark across 8 languages reveals significant gaps — and a path forward.

Automatic spellcheckers silently erase the very signal that typing-error biometric systems rely on. We benchmarked how bad the damage is — and what it means for system design.

Machine learning powers modern Android malware detection — but it is vulnerable to backdoor attacks. We propose a feature-selection defence that reduces this risk without sacrificing accuracy.

People make characteristic mistakes when they type. Can those mistakes serve as a biometric signature for authentication? This project built — and rigorously evaluated — a system that answers yes.