India proposes DEPA framework to unlock AI training data

The proposed architecture, termed DEPA for AI Training, establishes a techno-legal environment where public and private sectors can contribute datasets for machine learning. Unlike existing systems focused on individual-level inference, this framework is designed specifically for training models. It relies on four pillars: discovery platforms like AI Kosh, machine-readable electronic contracts, secure hardware enclaves, and differential privacy techniques to mask sensitive information.

Training workflows follow a zero-trust model where datasets are registered, encrypted, and processed within secure environments. Data providers, such as hospitals and banks, maintain control over their information, while consumers—including startups and researchers—access the data through automated, cryptographically signed agreements. The draft also incorporates accountability measures, including bias audits and immutable logging via an optional distributed ledger to ensure traceability.

Government officials frame this infrastructure as a vital economic engine, citing the potential for AI to contribute $500 billion to India’s GDP by 2030. Shri Abhay Karandikar, Secretary of the Department of Science and Technology, described the initiative as a response to the talent-data mismatch, positioning it as foundational national infrastructure rather than mere regulation. The proposal draws on global precedents, such as the Penn Medicine and Intel federated learning project, to demonstrate that high-performance AI models can be trained without exposing raw, sensitive records.