Dr David Grigorovich Khachaturov

PhD thesis

Technical and societal implications of machine learning security

My PhD focused on the security of machine learning systems, spanning technical attacks, defences, and governance. It investigated new vulnerabilities in modern AI, including simple physical attacks on image classifiers, audio-based jailbreaks against multimodal large language models, and data poisoning in language models, showing that larger models can be more susceptible to harmful behaviour. It also explored effective dimensionality as a new way to measure model robustness, finding that lower-complexity models are often more resistant to attack. On the defence side, the project identified weaknesses in LLM watermarking and developed Adversarial Suffix Filtering – a lightweight, model-agnostic method for blocking prompt injection attacks. Beyond technical solutions, it proposed a tiered anonymity policy framework for social media to help limit AI-generated misinformation while preserving pseudonymity for ordinary users. Overall, the project sought out to improve the security, robustness, and trustworthiness of ML systems in society.

After the PhD

After completing my PhD, I plan to continue working at the intersection of AI security, research, and education. I am currently the Lead AI Security Researcher at FrontierOne, where I work on sovereign AI security capabilities, and I expect to keep building practical security solutions for real-world AI systems. Alongside this, I am developing an education-technology company – studysesh® – focused on improving how students learn, and I hope to continue growing that work as well. I also intend to remain involved in academia through my roles at Cambridge, including teaching, supervision, and my Bye-Fellowship at St Catharine’s College. Overall, my goal is to combine academic insight, industrial impact, and educational innovation in everything I do.