Privacy-First AI: Google Cloud & MLCommons Tackle Brain Tumors

Medical AI is advancing fast, but there’s a swamp-sized problem: how do you test models on real patient data without exposing sensitive info? Google Cloud and MLCommons are wading in with a solution that keeps everything under lock and key—even from Google itself.

Their partnership, announced at Cloud Next, uses Confidential Computing to create a secure clean room for benchmarking AI models. It’s like a private pond where only the approved fish can swim, and no one else gets a peek.

The Challenge: Evaluating AI Without Seeing the Data

MLCommons launched MedPerf in 2023 to standardize medical AI evaluation. The platform uses federated evaluation to test models across institutions without centralizing data. But even federated learning has risks—model weights and data can leak.

Enter Google Cloud’s Confidential Space. This hardware-isolated Trusted Execution Environment (TEE) encrypts memory in-use and hardens the OS. Neither hospitals, researchers, nor Google can see the model code or patient data during evaluation.

Medical AI is compute-heavy, so the protection extends to GPUs. MedPerf runs on A3 machines with NVIDIA H100 GPUs, pairing Intel TDX on the CPU with NVIDIA Confidential Computing on the GPU. Before any data is released, cryptographic proof ensures only approved code runs on genuine hardware.

From Theory to Impact: Brain Tumor Research

This tech is already making waves in the Federated Tumor Segmentation (FeTS) initiative. Brain tumors like glioblastomas are rare, so no single hospital has enough data for high-accuracy AI. Plus, models that work at one site may fail at another due to demographic and equipment differences.

Working with researchers like Dr. Spyridon Bakas (Indiana University), Dr. Yury Velichko (Northwestern), and Dr. Amber Simpson (University of Alberta), MedPerf is validating AI on private brain MRI data worldwide. They’re spotting performance gaps—a model might be 95% accurate at one site but only 63% at another. This ensures that when AI reaches clinicians, it’s proven across a representative patient population.

Clinical Trust and Validation

Dr. Velichko says, “The future of medical AI lies in secure, scalable, and collaborative cloud environments.” Alexandros Karargyris, MedPerf lead, adds, “We have taken a major step toward a future where AI models can be rigorously tested on real patient data—without compromising privacy.”

The Future: Scaling Secure Breakthroughs

This collaboration marks a shift toward privacy by design in healthcare AI. By making secure data sharing easier, we’re clearing the path for faster, safer medical breakthroughs. If you’re a research institution or model developer, contact [email protected] or your Google Cloud account team.

For more on Google Cloud’s capabilities, check our Google Cloud review. And if you’re comparing AI models, our pricing comparison might help.

Original announcement published on Google Cloud.