Product introduction

The federated modeling service uses a joint computation model. This lets you develop and evaluate models without moving your raw data.

The federated modeling product is the visual platform for the federated modeling service. It helps federated model developers process data and build and evaluate models. Federated modeling integrates technologies such as federated learning, trusted execution environment (TEE), Secure Multi-party Computation (MPC), and differential privacy (DP). These technologies protect against differential attacks and secure the intermediate information of each participant.

Benefits

The federated modeling service has the following benefits:

• Secure

  To address potential data security issues during federated modeling, the product integrates multiple security solutions from cryptography, information theory, and encrypted hardware. This integration achieves a balance between security and performance. It enables privacy-preserving computation for joint modeling while adhering to various compliance policies.

• Easy to use

  The product provides interactive programming, high-level API encapsulation, and integration with common machine learning libraries. It offers out-of-the-box capabilities and a workflow experience similar to centralized data modeling. This lets you quickly begin model development and efficiently build models.

• High performance

  To address potential high latency and low bandwidth issues among multiple participants, the product features an optimized communication mechanism. For example, it uses model and gradient sparsification to reliably support large-scale, distributed model training. To handle heterogeneous data and computing power, it uses personalized client model algorithms to optimize and improve performance.

• Scalable

  This product shares the same core Resource Management platform and system framework as other products on the Ant Privacy-Preserving Computation Platform. You can combine product features as needed to create a complete and precise solution.

Scenarios

Multiple organizations may have data with similar structures but different content. Each organization wants to create a machine learning model using its own data. Using the federated modeling service, these organizations can jointly create a single machine learning model, called a federated model. This federated model performs better than a model trained on only one organization's data. If compliance policies require that raw data cannot leave its domain, federated modeling can help. It uses an approach where the model moves between participants, but the data stays in place. This allows each participant to benefit from a more effective machine learning model. The process secures the raw data and intermediate information, allowing all participants to share the benefits of their combined data value. The following diagram shows how multiple organizations can use the federated modeling service together.