The Defense Advanced Research Projects Agency (DARPA) has recently teamed with Slingshot Aerospace to create an AI powered system to provide better Space Domain Awareness (SDA). The new model, Agatha, is trained on over 60 years of simulated data that Slingshot has created to represent constellations in space. This training allows Agatha to quickly identify and alert on spacecrafts that are either malfunctioning or performing some type of nefarious action in space, a mission that will grow increasingly difficult as mega-constellations with thousands of satellites are launched over the next several years.
Agatha has already started to identify abnormal behavior of in orbit satellites, with a report being generated for a non-nominal orbital maneuver being generated for the Russian satellite Luch (Olymp) 2 that is in a GEO orbit. Over the course of several days in 2023, Agatha watched as Luch 2 maneuvered from its expected orbit to re-position itself near several other satellites at a different longitudinal position in GEO. Agatha was able to use the multitude of data sources available to it to quickly alert relevant parties to the initial maneuver by Luch 2 shortly after it occurred and was also able to indicate what orbital regime it may be transiting to based on its predictive modeling capabilities. This information enables other operators to take any necessary actions to avoid collisions of orbital assets or take other pre-emptive actions as required.
The Agatha model can automate the monitoring actions of individual operators, and essential task that needs to be developed before mega-constellations overwhelm operators, however it also creates risks for SDA moving forward. The Agatha model is only as good at the data used to train it and the new data used to make predictions, and therefore is at risk of a data poising attack.
Due to the lack of historical mega-constellation data, Slingshot Aerospace had to generate its own simulated test data to train the model to identify anomalies and nefarious actions. Though this gives them many decades more of data than would be available from real on-orbit data (with mega-constellations only being on orbit for ~5 years), the accuracy of the data being used must be heavily investigated. With 60 years of data, it is likely that many millions of data points were used to train the model. This overwhelming amount of data would create an easy opportunity for a threat actor to inject bad data points into the model, thereby training the model that an action they wish to perform should be seen as ‘normal’ and allow them to go un-noticed.
Additionally, while the model is deployed for use it draws from multiple data sources including Slingshots own Sensor Network as well as other owner/operator and SDA sources. While this allows for a more comprehensive look into what is occurring in space, it again creates risks of incorrect data being injected into the model to hide adversary actions. By modifying data sets being input to the model, an orbital maneuver could take place un-noticed. Additionally, if a bad actor gains an understanding of the data filter that is being used, they may be able to manipulate data inputs such that data of their choosing gets ignored by the model.
While the use of Agatha and other AI/ML models for SDA related actions is an innovative and necessary step forward, it carries with it risks in our understanding of space. Special attention needs to be placed in how these models are trained and what they are trained on to ensure accurate alerts and predictions are made. Data sources that are used in the models need to carry with them enhanced security and must put high levels of importance on data integrity to ensure accurate real-time data is fed into the models. With an explosion of mega constellations forecast in the years ahead, it is important that SDA stakeholders embrace the use of AI in operations while safeguarding the data used in these models.
https://www.slingshot.space/news/slingshot-darpa-agatha-ai
https://www.slingshot.space/news/russian-luch-olymp-2-satellite-approaching-multiple-geo-spacecraft