Challenge

Facial recognition (FR) technology is increasingly used in policing to enhance security and operational efficiency. However, testing these systems with real-world data raises significant ethical concerns, including privacy violations and potential biases.

Our client, a UK police authority, sought a solution that would allow for comprehensive testing of FR systems without relying on actual images of individuals. The objective was to develop a method to generate synthetic facial data that accurately represents diverse demographics, enabling controlled and repeatable testing scenarios.

Solution

Butterfly Data developed a synthetic face generation tool leveraging generative AI techniques. The solution utilised an open-source model as the foundation, incorporating a custom Python-based prompt generation engine to guide the creation of realistic, diverse mugshots.

Key features of the tool included:

• Ability to specify attributes such as age, ethnicity, and gender to create varied datasets
• Techniques to reduce stereotypical representations related to clothing styles, hairstyles, and other features
• Generation of faces under different lighting conditions and expressions to test system robustness

This approach enabled the creation of large-scale, diverse synthetic datasets tailored to specific testing requirements, facilitating ethical evaluation of FR systems.

Impact

The synthetic face generation tool provided the police authority with the capability to:

• Conduct comprehensive evaluations of FR systems without using real-world images, addressing privacy concerns
• Identify potential biases in FR algorithms by testing across a wide range of demographic variables
• Inform adjustments to FR systems, enhancing accuracy and fairness in real-world applications

The project not only advanced the ethical testing of FR technology but also laid the groundwork for future research into responsible AI deployment in policing contexts.