Google Cloud and MLCommons move MedPerf medical AI benchmarks into confidential computing

Google Cloud says it is working with MLCommons to run MedPerf medical AI evaluations inside confidential computing environments, a move aimed at letting researchers test models on sensitive clinical data without exposing patient records or proprietary model code.

The August 6 announcement focuses on a practical bottleneck in health AI: models need to be validated against diverse, real-world patient data, but hospitals and research institutions face strict privacy and governance limits on sharing that data. MedPerf, an MLCommons initiative launched in 2023, already supports federated evaluation so models can be benchmarked across institutions. Google Cloud's contribution is to place that process inside hardware-isolated trusted execution environments.

Why it matters

In Google's description, the setup uses Confidential Space and a hardened Confidential VM so that the hospital, other participants, and Google cannot inspect the model code or patient data while the evaluation runs. For compute-heavy medical imaging workloads, the environment extends protection to GPU-backed inference on Google Cloud A3 machines with NVIDIA H100 GPUs, pairing Intel TDX on the CPU side with NVIDIA Confidential Computing on the GPU.

Before data is released into a workload, the system provides cryptographic attestation that the approved code is running on genuine confidential computing hardware in a hardened environment. That mechanism is important because medical AI validation is not only a performance problem; it is also a trust problem for hospitals, model developers, regulators, and clinicians.

Brain tumor research is the first showcase

Google points to the Federated Tumor Segmentation initiative as an early use case. Brain tumors such as glioblastomas are rare, so one hospital may not have enough cases to evaluate a model reliably. Differences in scanners, acquisition techniques, demographics, and local clinical practice can also cause a model that performs well at one site to degrade elsewhere.

The company says MedPerf on Google Cloud is being used to validate AI models on private brain MRI data from multiple institutions and to find performance gaps across sites. Google gives one example in which a model could score 95 percent at one site and 63 percent at another, underscoring why single-site validation can be misleading.

The announcement does not make a regulatory approval claim or say that any specific clinical product is cleared for deployment. Its significance is narrower but still meaningful: it gives medical AI teams a more secure path for independent, multi-site benchmarking, which could make future model evaluations more representative without requiring hospitals to pool raw patient data.

Sources

• cloud.google.com
• mlcommons.org

Cover photo by cottonbro studio on Pexels, used under the Pexels License.