Machine Unlearning for Privacy-Preserving Cross-Modal Retrieval Systems (Ref: CO/GC-SF6/2025)

This PhD project addresses the critical challenge of selective information removal and privacy preservation within cross-modal retrieval systems. Contemporary cross-modal architectures require effective mechanisms to selectively forget specific data points across multiple modalities whilst maintaining retrieval performance – a fundamental requirement in privacy-conscious and adaptable multimedia retrieval systems.

Background:

Cross-modal retrieval has revolutionised traditional multimedia search systems through deep learning implementations. Whilst current approaches, including vision-language pre-training models (CLIP, ALIGN), Transformer-based architectures, and GNN-based frameworks, demonstrate impressive capabilities in aligning different modalities, they lack robust mechanisms for selective information removal. This limitation becomes particularly crucial when handling privacy requests, outdated content, or the need to correct misaligned multi-modal pairs.

Machine unlearning in cross-modal retrieval presents unique challenges beyond traditional unlearning scenarios. The complex relationships between different modalities (text, images, video, audio) mean that removing specific information can have cascading effects across modalities and impact overall retrieval performance. Additionally, simply deleting information without considering cross-modal relationships can lead to semantic gaps and inconsistencies in retrieval results, potentially compromising the system’s effectiveness.

Current unlearning approaches, primarily designed for single-modality tasks, prove inadequate for cross-modal systems due to the unique challenges of maintaining semantic alignment across modalities whilst selectively removing information. Recent research indicates promising directions in developing specialised unlearning strategies for multi-modal systems, but fundamental challenges remain in balancing effective forgetting with maintained cross-modal retrieval performance.

Aim: 

This project aims to advance the field of cross-modal retrieval by addressing crucial privacy and adaptability requirements, with direct applications in privacy-preserving multimedia search engines, cross-modal recommendation systems, secure multi-modal information retrieval, and adaptable visual-semantic platforms.

Research Objectives:

1. Develop novel machine unlearning methodologies specifically designed for cross-modal retrieval systems, focusing on selective forgetting across multiple modalities whilst preserving semantic alignment and maintaining retrieval performance on retained information across different modalities.
2. Create comprehensive evaluation frameworks for cross-modal unlearning effectiveness, incorporating metrics for cross-modal forgetting quality, semantic alignment preservation, retention performance, and privacy protection measurement.
3. Design efficient compensation mechanisms that enable the replacement of forgotten information across modalities with suitable alternatives, maintaining semantic consistency and cross-modal relationships whilst ensuring privacy and retrieval accuracy.
4. Investigate the impact of different cross-modal architectural choices and multi-modal data characteristics on unlearning effectiveness, developing guidelines for unlearning-aware cross-modal retrieval system design.

94% of Loughborough’s research impact is rated world-leading or internationally excellent. REF 2021

Start date

July 2025, October 2025, January 2026, April 2026

Supervisors

• Primary supervisor: Professor Georgina Cosma

Entry requirements

Applicants should have, or expect to achieve, at least a 2:1 Honours degree (or equivalent) in computer science or a related subject. A relevant Master’s degree and/or experience in one or more of the following will be an advantage: artificial intelligence, information sciences, mathematics with experience in programming.

English language requirements

Applicants must meet the minimum English language requirements. Further details are available on the International website.

Tuition fees for 2025-26 entry

• UK fee – To be confirmed Full-time degree per annum
• International fee – £28,600 Full-time degree per annum

How to apply

All applications should be made online. Under programme name, select Computer Science. Please quote the advertised reference number: CO/GC-SF6/2025 in your application.

To avoid delays in processing your application, please ensure that you submit a CV and the minimum supporting documents.

The following selection criteria will be used by academic schools to help them make a decision on your application. Please note that this criteria is used for both funded and self-funded projects. 

Please note, applications for this project are considered on an ongoing basis once submitted and the project may be withdrawn prior to the application deadline, if a suitable candidate is chosen for the project.

Apply now