A team of scientists from the Warsaw University of Technology participates in the PINEBERRY project, implemented with the KP Labs team for the European Space Agency. Its aim is to develop procedures, catalogues of solutions and examples of solutions for security and explanation of AI models used in space missions. The work of the MI2.AI from WUT is led by Professor Przemysław Biecek.

The work is divided into two phases. The result of the first phase – development of procedures and catalogue of solutions – are two documents describing how to build safe and explainable AI models: ”Security Risks for AI Applications in Space Operations” and “XAI Techniques for AI Applications in Space Operations”. In addition to the technical version, there is also a popular science version in the form of a comic that presents key challenges in a more user-friendly way. Explainability and transparency of models is important for operators of space missions using AI models to have more trust in them and so use them more willingly.

- These case studies make up a coherent story which is the basis for a comic book promoting the applications of eXplainableAI in Space. It may seem that a comic book is not a significant scientific achievement but in this project it is very important to communicate complicated methods in a way clear for space mission operators, so very busy people. A lighter form was very useful for that purpose – explains Prof. Przemysław Biecek, leading the work of WUT scientists.

The result of the second phase will be 5 demonstrators of AI securities for space missions for telemetry, images and text data.

Beginning of ”space” cooperation

The PINEBERRY project is implemented in cooperation with the European Space Agency (ESA), European Space Operations Centre (ESOC), European Space Research and Technology Centre (ESTEC) and KP Labs – a Polish space company. 

- The European Space Agency opened admissions for research teams that would develop the above mentioned catalogues. We were invited to the PINEBERRY project by KP Labs, which specialises in developing AI tools for space applications. Our team MI2.AI from WUT supports KP Labs in terms of AI explainability and security. We have written together a grant application which won a competition and was awarded funding. The project has been implemented since March 2024 – says prof. Przemysław Biecek.

The role of prof. Biecek is to manage the work of WUT scientists, as well as propose solutions, mainly pointing to existing ones or developing new ones, in the field of explainable AI.

- An example result I am very happy with is development of a series of case studies on how to include explainability into analysis of various types of AI models in space, starting with predictions of energy demand for a Mars rover, to detecting space debris – emphasises Prof. Biecek. – For us a challenge was the versatility of AI uses. There isn’t one model that should be studied, explained and secured. AI models for computer vision, text, timeframes are a very dynamic and quickly developing area of research so there are lots of models like that in space missions.

Apart from prof. Biecek, WUT academics who participate in the project are from the Faculty of Mathematics and Information Science (MiNI) and the Faculty of Electronics and Information Technology (EiTI). Professor Artur Janicki from EiTI is an AI security expert and three talented students and graduates of MiNI: Piotr Wilczyński, Agata Kaczmarek and Dawid Płudowski, are developing sample programmes that use the techniques of AI explainability and security in the proposed solutions. The AI security team developed a catalogue of 9 most frequent susceptibilities of AI models, weaknesses that can be abused by an unauthorized person. Then for each of them they developed security methods – that is ways of securing the model against attacks.

The work is being done as the project ends in April and then the developed solutions are planned to be published. The beneficiaries of the project are first of all space mission operators.

- We will have access to safer, better studied and more transparent AI models. In the next step these will be hundreds of companies and institutions that use data collected in space missions. For example, one solution is connected with analysis of soil parameters in terms of applications in agriculture on the basis of hyperspectral images from satellites. Better reading of soil parameters is support for smart agriculture - explains Prof. Biecek.

Importantly, the MI2.AI WUT team is working on improving AI models in terms of security and explainability in various fields. In addition to cooperation with KP Labs in the space sector, the team is also working on health applications in radiology and oncology. As prof. Biecek points out, our scientists are open to cooperation with other partners who need to upgrade their models in terms of security or explainability.