Submitted Grant Application to Alan Turing Institute's Defence & Security Program

Update (23 September 2024): While our proposal was not selected for funding in this round, we received notification that the overall standard of applications was very high. This experience has been valuable for developing grant writing skills and refining our research ideas. We plan to incorporate any feedback received and continue pursuing funding opportunities for this important research direction.

Under the guidance of my PhD supervisors Dr. Budi Arief and Prof. Shujun Li, I submitted a grant application to the Alan Turing Institute’s Defence & Security Grand Challenge: AI Security program. The proposed project, titled “Investigating the Impact of Fine-tuning Open-source Models on Cyber Skills Performance,” addresses the critical need to understand and mitigate security risks in fine-tuned AI models.

The proposal responds to a specific challenge from the Alan Turing Institute regarding the investigation of how fine-tuning open-source models impacts cyber capabilities. This research is particularly timely given the increasing accessibility of open-source AI models and their potential dual-use nature in cyber security applications.

If funded, the research aims to:

• Investigate the security implications of fine-tuning open-source AI models
• Assess how fine-tuning affects model behavior and potential misuse scenarios
• Develop robust methodologies for secure model adaptation
• Create comprehensive guidelines for maintaining model safety
• Establish best practices for the responsible fine-tuning of AI models

The proposed methodology encompasses a systematic 6-month research plan:

1. Model Selection and Dataset Curation
   • Identification of suitable open-source AI models
   • Collection and preparation of diverse cyber security datasets
   • Development of evaluation metrics
2. Fine-tuning Implementation
   • Implementation of various fine-tuning techniques
   • Assessment of model behavior changes
   • Documentation of performance variations
3. Security Evaluation
   • Comprehensive vulnerability analysis
   • Testing against various attack scenarios
   • Assessment of safeguard effectiveness
4. Mitigation Strategy Development
   • Design of security enhancement measures
   • Testing and validation of protective mechanisms
   • Documentation of successful approaches
5. Guidelines and Best Practices
   • Synthesis of research findings
   • Development of practical recommendations
   • Creation of implementation frameworks

The proposal was submitted on September 1, 2024, with an anticipated start date of October 2024 if successful. The research would be conducted at the Institute of Cyber Security for Society (iCSS) at the University of Kent, leveraging the institute’s extensive expertise in cyber security and AI.

Special acknowledgment to my supervisors for their exceptional guidance throughout the application process, helping shape both the technical aspects and broader impact of the proposed research.

Project updates will be shared here if the grant application is successful.