Master thesis in the area of Privacy-Preserving Machine Learning: Zero-Knowledge Proofs for Fairness Verification in Machine Learning

Goal

Develop and evaluate cryptographic methods that allow a machine-learning model owner to prove that a model satisfies selected fairness criteria without revealing sensitive information about the underlying dataset, model outputs, or computed fairness statistics.

Background

Fairness in machine-learning systems is increasingly important in applications where automated decisions may affect different demographic or protected groups. Verifying fairness typically requires access to sensitive data, predictions, or aggregate statistics, which may not be desirable or possible due to privacy or confidentiality requirements.

Zero-knowledge proofs provide a way to demonstrate that a computation has been performed correctly without revealing the private inputs used in that computation. Applying these techniques to fairness verification could allow an external party to verify that a model satisfies a chosen fairness criterion while keeping sensitive evaluation information private.

Tasks

The student will:

• Study common fairness criteria for machine-learning systems, starting from statistical parity.
• Identify the information and computations required to evaluate selected fairness properties.
• Investigate cryptographic commitments and zero-knowledge proof techniques suitable for privately proving these computations.
• Desing a protocol in which relevant values, such as group sizes and prediction counts, remain hidden while their relationships can be verified.
• Evaluate the proposed approach in terms of proof size, computation time, and scalability.

Student Profile

• Interest in cybersecurity, cryptography, privacy or trustworthy machine learning.
• Basic programming experience.
• Basic knowledge of machine learning and cryptography is useful, but advanced knowledge of zero-knowledge proofs is not required.
• Motivation to work on a research-oriented problem combining machine learning and modern cryptography.