Medical AI faces a challenge: models must be tested on diverse real-world patient data, but doing so risks exposing sensitive information and proprietary model code. Google Cloud and MLCommons address this through the MedPerf initiative, an open-source benchmarking platform that uses federated evaluation and Confidential Computing. By running inside hardware-isolated Trusted Execution Environments (TEEs) on Google Cloud’s A3 machines with NVIDIA H100 GPUs, neither hospitals, research institutions, nor Google can see model weights or patient data during inference. Cryptographic attestation verifies that only approved code runs on genuine Confidential Computing hardware.

The approach is already seeing critical use in the Federated Tumor Segmentation (FeTS) initiative for brain tumor research. Because glioblastomas are rare, single hospitals lack sufficient data; models that achieve 95% accuracy at one site can drop to 63% at another due to demographic and equipment differences. MedPerf on Google Cloud validates AI models on private brain MRI data from multiple sites, identifying performance gaps before deployment. Researchers like Dr. Yury Velichko (Northwestern University) emphasize that moving from lab settings to production-ready infrastructure provides unique real-world validation.

The collaboration represents a shift toward privacy by design in healthcare AI. The article does not discuss limitations or scalability boundaries. Interested institutions can contact MLCommons or their Google Cloud account team.