PhD Studentship: Robust, Certified, and Scalable Federated Machine Unlearning for Privacy-Preserving AI

This PhD studentship at the University of Surrey focuses on advancing federated machine unlearning to address the critical challenge of data erasure in privacy-sensitive domains like healthcare and finance. The project aims to create scalable, certified, and robust methods for removing specific data influences from trained models without compromising performance or privacy. You will explore cutting-edge topics such as certified unlearning with formal guarantees, adversarial robustness, and unlearning for foundation models, contributing to trustworthy AI and regulatory compliance. The work sits at the intersection of privacy-preserving machine learning, distributed systems, and AI safety, with real-world implications for federated systems deployment.

• Develop next-generation algorithms for federated machine unlearning to ensure verifiable data erasure in privacy-preserving AI systems.
• Design methods with provable erasure bounds, connecting to differential privacy and statistical divergence frameworks.
• Investigate robustness of unlearning protocols against adversarial relearning attacks, including fine-tuning and jailbreak-style probing.
• Build federated-specific benchmarks and auditing tools to evaluate and verify unlearning methods.
• Extend federated unlearning techniques to large pretrained models, balancing erasure guarantees with model utility.
• Explore how federated unlearning can contribute to removing hazardous or harmful knowledge from collaboratively trained models.
• Collaborate with supervisors to publish findings in top-tier conferences and journals within the field of privacy-preserving AI.