How the orchestration layer securely enables AI insights in a federated learning framework
-
The CAIA orchestration layer acts as a conductor for the federated learning process, facilitating collaboration between cancer centers without ever accessing sensitive patient data.
It leverages NVIDIA FLARE to provide the essential framework and code for secure communication, with cancer centers actively initiating a secure connection and executing local tasks after receiving code from the orchestrator.
The layer has two key functions: coordinating the network by tracking connected centers and ensuring they are ready, and aggregating the local model updates (summarized insights) sent back to create an updated global model.
This setup ensures patient data remains localized and protected behind each center's firewall while collective, de-identified insights are securely combined to create more accurate AI models for cancer research.
In our explainer blog post on federated learning, we elaborated on the technical setup that enables this AI training approach. One of the most crucial aspects of this setup is the central orchestration layer or node that facilitates collaboration between the participating cancer centers without ever accessing sensitive patient data.
As we previously explained, the orchestration layer acts as a conductor for a musical performance. This layer only sees the model updates or summary insights that the AI models learn locally from each cancer center.
In this post, we’ll take a closer look at the technology that underlies this secure orchestration layer.
How the orchestration layer works
The CAIA platform leverages NVIDIA FLARE (which stands for Federated Learning Application Run-time Environment), an open-source library that provides a framework to control the communication between components of the federated network. More specifically, FLARE provides the essential structure and code that allows the orchestration layer and the cancer centers to connect and communicate with each other.
View of how the orchestration layer works
Each of the cancer centers receives a piece of the FLARE-based software package that polls the orchestrator for tasks and executes those tasks locally. Each cancer center must actively connect to the central orchestration layer. The orchestration layer can neither access the cancer centers’ data nor independently initiate contact; it relies on the cancer centers to establish a secure connection.
The orchestration layer has two primary but inter-related functions:
As a coordinator: By tracking which cancer center is connected and active, ensuring that the cancer center is operating as expected and ready to receive model weights and summaries from CAIA.
As an aggregator: It takes the updated model weights (summarized insights) sent back from the cancer centers and combines them to create a stronger, more generalized global model.
Why is the orchestration layer important?
Much like the conductor guiding a musical performance, CAIA’s orchestration layer is essential to the entire federated learning process:
The orchestration layer is responsible for model distribution. It sends the initial AI model and instructions to each of the cancer centers
Each cancer center then trains the AI model using its own local, secure data. The updated AI model only contains summarized learning or insights. No actual patient data or private information is ever shared. These summaries or insights then travel back to the orchestration layer to update the global AI model.
The orchestration layer combines the learnings from each cancer center one at a time to create a more powerful, global model. This updated, more intelligent model is then sent back to the centers and the cycle begins again. Each iteration makes the model smarter and more accurate, all while the patient records themselves remain safely behind each institution’s firewalls.
The CAIA orchestration layer makes privacy-preserving cancer research possible at scale. By serving as both a coordinator and aggregator, it oversees the repeated process of federated learning. This setup ensures that patient data remains localized and protected at the cancer centers, while the collective, de-identified insights from diverse centers are securely combined to create more equitable models.
If you’d like to learn more about CAIA, subscribe to our newsletter for updates and follow us on LinkedIn and X.
