At the lower end of the risk scale, individual speculators manipulate weather stations for personal gain, as seen at CDG Airport. A significant concern arises when groups of traders coordinate to skew renewable energy output forecasts, leading to volatility in wholesale power prices and causing losses for other stakeholders. On the extreme end, state actors or jammers can interfere with broadcast stations, triggering false alarms in early warning systems or silencing them when alerts are necessary. The spectrum of risks ranges from fraud to compromised disaster preparedness and national security issues, escalating alarmingly.
Given that there are continuous incentives—be they financial or otherwise—for manipulating observational data, adversaries will always seek new opportunities. Our objective is to stay one step ahead. Here are three effective strategies.
1. Monitor Broadcast Stations Continuously. Implementing robust data quality control measures is essential, including security for stations, anomaly detection, correction mechanisms, and human oversight. Weather stations need constant monitoring to prevent tampering. Additionally, swift data homogenization methods are crucial for cleaning weather records and gaining real-time insights. This becomes increasingly vital as AI systems utilize this data for immediate decision-making. Human oversight is indispensable for flagging suspicious data and model outcomes—after all, it was human observation that uncovered the manipulation at CDG Airport.
2. Secure Your Data to Safeguard AI. Establish comprehensive data security measures across the entire AI pipeline. Tools for AI explainability and adversarial robustness can elucidate the underlying data and outputs of your AI models, helping you identify pertinent data or model issues while bolstering resilience against potential attacks.
3. Ensure Continuous Accountability Across the Chain. Observational data passes through various entities, including operators at observatories, the National Weather Service, and forecast centers transforming data into actionable forecasts. Data integrity cannot be safeguarded by any single entity; each must protect its own link in the chain. Furthermore, any anomalies must be communicated across the entire chain—from station operators to the entities acting on the forecasts.
We are fortunate to have insight into the situation at CDG Airport, but it serves as a critical wake-up call. As the reliance on observational data in weather prediction intensifies, we must adapt to evolving threats. This entails fortifying our data and models by enhancing existing oversight and accountability frameworks and fostering better collaboration among all involved partners.
This editorial was authored by:
- Monique Kuglitsch — Innovation Manager at the Fraunhofer Heinrich Hertz Institute and Chair of the United Nations Global Initiative for Natural Disasters through AI Solutions
- Jesper Dramsch — Machine Learning Scientist at the European Center for Medium-Range Weather Forecasts (ECMWF). He is involved in developing ECMWF’s data-driven weather forecasting model, AIFS (Artificial Intelligence Forecasting System).
- Franz G. Kuglitsch — Climate Scientist and Executive Director of the International Union of Geodesy and Geophysics (IUGG) at the GFZ Helmholtz Center for Geosciences in Potsdam.
- Andrea Toretti — Senior Researcher at the European Commission Joint Research Center (JRC), coordinating European and global drought observatories within the Copernicus Emergency Management Service.
Source: www.technologyreview.com