India's AI framework lets models enter secure enclaves for local training without exporting raw data

India’s Ministry of Electronics and Information Technology has proposed a new regulatory framework for artificial intelligence that keeps training data within secure domestic boundaries. The scheme, known as DEPA for AI Training, seeks to regulate and secure the flow of information to create AI models trained on local data.

The evidence

According to The Register AI/ML, the proposal centres on the principle that the model travels to the data rather than the data travelling to the model.reported Professor Ajay Kumar Sood, Principal Scientific Adviser to the Government of India, described this approach as the core mechanism of the Data Empowerment and Protection Architecture. The framework envisions Training Data Providers, such as hospitals and banks, sharing anonymised and encrypted data with a central Data Exchange Platform.

The decision

Training Data Consumers would access this stored data to train their models without extracting the raw information. MeitY’s documentation states that these exchanges will operate hardware-based Secure Data Environments where data remains encrypted and isolated even during processing. The system also enforces usage rules defined by the data providers, ensuring that consumers adhere to specific constraints on how the data is utilised.

The takeaway

The framework describes the systematic application of differential privacy techniques to stop models from learning or revealing individual-specific information. All activities within the secure environments are immutably logged for audit purposes, which the framework states ensures the process is both secure and transparent. This structure attempts to balance the need for AI innovation with strict data sovereignty and privacy controls.

Building on this

Developers can build applications that train models on sensitive datasets without ever exposing raw records to external servers. The documented interface involves connecting to a Data Exchange Platform that hosts hardware-based secure enclaves. Builders must design models capable of travelling to these enclaves, meaning training logic must be deployable within restricted, isolated environments. Limits include strict adherence to provider-defined usage rules and the requirement to use differential privacy techniques. The architecture supports sectors like healthcare and finance, where data providers include hospitals and banks. Developers must account for immutable logging of all activities, which affects debugging and auditing workflows. The model must not learn individual-specific information, constraining the granularity of training data. This setup allows for local AI development while keeping data encrypted during processing. Builders cannot assume open access to data; they must negotiate terms with Training Data Providers.