LEAKPRO: Leakage Profiling and Risk Oversight for Machine Learning Models

Important results from the project

The project goals were achieved through the development of LeakPro, an open-source software framework for measuring information leakage in AI models, synthetic data, and federated learning. It supports multiple modalities and model architectures. LeakPro is part of Scaleout’s offering and has been used by Syndata in collaboration with the IMY. The project resulted in four scientific publications, collaborations with the public sector and OECD, and the training of seven master’s students.

Expected long term effects

The project is expected to contribute to safer and more reliable use of AI through methods to identify and measure leaks of sensitive information. LeakPro creates long-term benefits as an open source platform for research, industry and the public sector, and strengthens the possibilities to build AI systems that meet requirements on integrity, security and regulatory compliance. The project also contributes to continued competence building and international collaborations in the area.

Approach and implementation

The project was carried out in close collaboration between academia, the public sector and industry. AI Sweden and RISE were responsible for LeakPro´s basic architecture and attacks on trained models. Scaleout was responsible for federated learning and Syndata for synthetic data. AstraZeneca, Sahlgrenska and Region Halland contributed with use cases and requirements. The work was carried out iteratively with joint development and integration into the open source-based platform LeakPro.

External links

LeakPro Software