Introduction
Privacy-preserving technologies are increasingly embedded into AI systems to ensure legal compliance and build public trust.
wo key mechanisms are:
- Differential Privacy (DP), which ensures individual training data cannot be inferred from model outputs;
- Machine Unlearning (MU), which enables AI models to remove the influence of specific data points after training, as required under regulations like the GDPR.
While both methods protect user privacy, they pose a significant challenge: there is no standard, efficient way to verify that these protections have actually been applied. Service providers may need to prove to regulators, auditors, or even end users that Differential Privacy was correctly enforced during training, or that Machine Unlearning was properly executed in response to a data deletion request. Without such verification, compliance claims remain untrustworthy and unverifiable.
Cryptographic techniques such as Zero-Knowledge Proofs (ZKPs) offer a powerful way to enable this verifiability without revealing sensitive model details or data. However, they are computationally intensive and energy-hungry, especially when applied to modern AI models at scale or in real-time applications — making their practical deployment a significant sustainability concern.
This project aims to explore how hardware-software co-design can enable energy-efficient verification of privacy-preserving AI mechanisms. Specifically, the student will:
Develop verifiable protocols for both DP (during training) and MU (during model updates and deletion requests);
Design and evaluate Zero-Knowledge Proof systems that certify the correct application of DP noise or MU adjustments;
Implement computationally intensive components (e.g., matrix operations, commitments, modular arithmetic) on Field-Programmable Gate Arrays (FPGAs);
Analyse the trade-offs between privacy guarantees, computational cost, accuracy, and energy consumption.
A key innovation is treating verifiability as a first-class sustainability concern. As privacy regulations become stricter, large-scale systems will need to provide auditable proof of compliance. Doing so at low energy cost is critical for the long-term viability of responsible AI.