The Ministry of Electronics and Information Technology (MeitY) has launched a handbook outlining a framework for sharing data securely for artificial intelligence (AI) training, as India seeks to expand access to datasets without compromising privacy, commercial confidentiality or institutional control.
The Handbook on Data Empowerment and Protection Architecture (DEPA) for AI Training was unveiled by MeitY Secretary S Krishnan at a roundtable in Bengaluru on October 31. The initiative was developed with support from the iSPIRT Foundation.
The framework addresses a problem faced by AI developers and researchers. Large datasets held by hospitals, banks, government departments and other institutions could help train more capable models, but concerns over privacy, confidentiality, intellectual property and data ownership often restrict access.
DEPA proposes a common system for data providers and AI developers to collaborate under defined permissions and security controls. The ministry's handbook describes the document as a resource that can evolve alongside the framework, with recommendations invited from institutions and individuals.
How DEPA proposes to protect data
The framework separates the institutions that hold data from those seeking to use it for AI development. Hospitals, banks and public agencies could act as Training Data Providers, while startups, researchers and enterprises could participate as Training Data Consumers.
Under the proposed arrangement, data providers would retain control over their records. Data-sharing agreements would specify permitted uses, while secure computing environments would allow AI models to be trained and tested without exposing raw data directly to developers.
The framework identifies four key components: data exchange platforms, electronic contracts, secure data environments and differential privacy. Data exchange platforms would connect providers and users, while electronic contracts would record permissions and usage conditions. Secure environments would host computing tasks, and differential privacy would help limit the exposure of information about individual records during analysis and model training.
The model is designed around a principle described by Principal Scientific Adviser Ajay Kumar Sood at the launch: “Under DEPA, the model travels to the data, rather than the data travelling to the model.”
The approach could allow institutions to collaborate on AI projects without transferring entire databases to outside developers. The handbook also proposes data provenance mechanisms to track where data originated, what permissions apply to it and how it is used during training
Proposed applications across sectors
The handbook identifies healthcare, financial services, education and geospatial data as areas where secure data collaboration could support AI development.
In healthcare, researchers could use data from multiple institutions to develop diagnostic tools, support drug discovery and improve public health research. In finance, access to broader datasets could help develop fraud detection and credit-risk models. In education, the framework could support personalised learning systems and portable skill credentials.
These are proposed applications rather than confirmed outcomes of the launch. The handbook argues that combining data from different institutions could help AI systems account for local conditions and populations, rather than relying on limited or predominantly foreign datasets.
The document also identifies potential economic gains. It estimates that operationalising the framework across sectors could help unlock $500 billion in AI-driven value for India's gross domestic product by 2030. This is a projection presented in the handbook, not a government estimate of benefits already realised.
Privacy risks and the need for oversight
Removing names and other direct identifiers does not always make a dataset anonymous. The handbook notes that information can sometimes be linked with other datasets to identify individuals, even after obvious identifiers have been removed.
To address this risk, the framework proposes anonymisation, encryption, secure computing environments and privacy-enhancing techniques. It also envisages continuous monitoring, security audits, access logs and measures to detect unauthorised activity. The document proposes an oversight body to set standards, register eligible participants, supervise data-sharing procedures and help resolve disputes.
The proposed safeguards will need to be tested in practice. Their effectiveness will depend on how institutions classify data, enforce permissions, monitor access and respond to misuse. Privacy-enhancing techniques can reduce risks, but their use does not automatically guarantee that every dataset or AI model is safe.
The handbook also raises questions about the interests of data providers and contributors. It proposes exploring incentives such as royalties or access fees to encourage institutions to make datasets available. The framework would need to establish clear terms on permitted use, accountability and the distribution of benefits generated from data-driven products.
How the framework relates to India's data protection regime
DEPA builds on India's wider digital public infrastructure and draws on the country's experience with consent-based data sharing, including the Account Aggregator system in financial services. The handbook distinguishes this earlier model, which supports authorised access to data for services such as lending, from DEPA Training's proposed arrangements for building and testing AI models.
The framework is intended to work alongside existing laws, including the Digital Personal Data Protection Act, 2023, and sector-specific regulations. It does not replace those legal obligations. The handbook also describes DEPA as a techno-legal framework, combining enforceable agreements with technical safeguards rather than relying on policy statements alone.
MeitY Secretary S Krishnan said at the launch, “India’s AI ambitions must be matched by the ability to use data responsibly and securely.”
The ministry said it would take the framework forward to support AI innovation while protecting privacy, commercial confidentiality and institutional control. The next test will be whether the proposed architecture can be implemented across institutions with different technical capabilities, data-sharing policies and regulatory obligations.
