Privacy-friendly AI predictions for shared sensitive data

Two companies want to team up. One holds private data shaped like a network: hospitals see how diseases spread, banks track transaction patterns. The other has a secret AI model that makes sense of such data to predict risks or trends. But neither can share their secrets directly—client privacy rules out raw graphs, while business owners don’t want to expose their model’s inner workings. A new approach called MAPP steps in as a kind of interpreter that keeps everyone’s sensitive parts locked away. Instead of forcing every AI shape into one rigid security protocol, MAPP builds small, fast “stand-in” models that learn to act like the real thing without seeing the real data. These lighter models run the same predictions but need less computing power during the risky live phase. By focusing on what matters most—patterns in the data—the framework shrinks the heavy math that normally happens while clients wait online. Most of that math moves to a safer offline stage when no one is watching closely.Real tests across seven different network datasets and four AI model types show that most lightweight stand-ins give results just as good as the originals. In some cases they even do slightly better. The real wins come after setup: partners can finish the job up to fourteen times faster online and cut data traffic by five to seven times. That speed gain matters when hundreds or thousands of users need answers without long waits. Behind the scenes, MAPP notices something most security tools miss: real-world networks are never dense. A hospital’s patient contacts list or a city’s bus routes contain lots of missing or zero links. By designing the protocol to skip heavy calculations on these empty spots, the system avoids a lot of unnecessary work. It’s like skipping empty pages when reading a book—you still get the story, just faster.