MeitY unveils DEPA handbook to enable trusted data sharing for AI training

DEPA stands for Data Empowerment and Protection Architecture. It is a framework for trusted data sharing during AI training. Its purpose is to help institutions use diverse, high-quality data while protecting privacy, confidentiality, intellectual property, and institutional control.

MeitY's handbook combines contractual permissions with controlled computing environments and privacy-enhancing safeguards. It also includes data provenance, so participants can trace data origins, permissions, and use throughout the training process. Under the model described by Principal Scientific Adviser Ajay Kumar Sood, the AI model travels to the data instead of the data travelling to the model.

MeitY launched the handbook with support from the iSPIRT Foundation. IT Secretary S Krishnan unveiled it at a Bengaluru roundtable attended by government, industry, academic, and research representatives. MeitY said it is committed to taking the framework forward as AI becomes more important to economic and social development.

In DEPA's approach, an AI model is brought into the environment where the data already resides. The underlying data does not need to be copied or transferred to another institution. This matters because privacy, confidentiality, intellectual property, and institutional-control concerns often restrict ordinary data sharing.

For example, several institutions could allow a model to train against their separately held datasets inside controlled computing environments. The model can learn from those permitted data sources, while the institutions keep the original information under their control. Contractual permissions define what is allowed, and privacy-enhancing safeguards help protect the data during the process.

Ajay Kumar Sood explained this principle at the Bengaluru roundtable. The article says it enables multi-party and cross-border collaboration, but does not describe a specific technical implementation or name a number of participating institutions or countries. MeitY presents the approach as part of responsible and secure AI development.

Contractual permissions specify who may use data, for what purpose, and under which conditions. Controlled computing environments keep the data within managed settings while an AI model works with it. Privacy-enhancing safeguards add protection against inappropriate exposure during the training process.

A research institution, for example, could permit an AI model to use selected data for a defined training project. The data could remain in the institution's controlled environment rather than being sent to the research partner. The agreement would set the permitted use, while technical controls and privacy safeguards would limit access and reduce risks. DEPA also adds provenance tracking to support oversight.

The handbook presents these measures as a combined techno-legal framework. The article does not identify particular contracts, software, or privacy technologies. It says the framework addresses concerns about privacy, confidentiality, intellectual property, and institutional control, helping institutions collaborate while retaining control of their data.

Data provenance is the record of where training data came from, which permissions govern it, and how it is used. In DEPA, that record continues throughout the AI training process. This matters because institutions need to understand whether data is being used lawfully, responsibly, and for the purpose they approved.

For example, a dataset contributed by one institution could carry information about its origin and permitted use. As a model trains with that data, the process can record how the dataset was used. If questions arise, collaborators have a trail linking the data to its source, permissions, and training activity. This works alongside contracts, controlled computing, and privacy safeguards.

The handbook says provenance supports accountability and transparency. It does not provide a particular software tool, record format, or audit procedure. Its role is to make data use traceable, helping institutions collaborate on AI training while preserving oversight and institutional control.

The article describes DEPA as supporting multi-party and cross-border collaboration. It does not state a maximum or fixed number of institutions, and it does not give a number of countries. Therefore, the framework's stated reach is broad in principle rather than quantified in the announcement.

For example, institutions in different countries could allow one AI model to work with data held in their respective controlled environments. Contractual permissions would define acceptable use, while privacy safeguards and provenance tracking would support protection and accountability. The model would travel to the data, reducing the need to move underlying datasets across institutional or national boundaries.

Cross-border participation matters because AI development depends on diverse, high-quality data, and collaboration can bring together institutions with different datasets and expertise. Sood described DEPA as India's contribution to global science diplomacy. The article does not report a completed international project or identify participating countries.

AI research and development become more constrained when institutions cannot share data. The article says access to diverse, high-quality data is critical as AI becomes central to economic and social development. If privacy, confidentiality, intellectual-property, or institutional-control concerns block access, researchers may have fewer datasets available for training.

For example, organizations may hold useful data but decline to provide it to a research partner because they cannot protect confidential information or retain control. That can prevent a joint project from using the data at all. DEPA addresses this barrier by combining contractual permissions, controlled computing environments, privacy-enhancing safeguards, and provenance tracking.

The article does not quantify delays, costs, or lost research opportunities caused by restricted sharing. It says these concerns often limit data sharing, which makes responsible data access an important condition for AI progress. MeitY's handbook is intended to help institutions collaborate more securely while keeping control of their data.

AI model training is the process of exposing a model to data so it can learn patterns and relationships relevant to a task. After training, the model uses what it learned to produce outputs or support decisions. The article does not explain training mechanics in detail, but it says access to diverse, high-quality data is critical as AI grows in importance.

Large datasets provide more examples for learning, while diverse datasets represent a wider range of subjects, conditions, and perspectives. For example, institutions could collaborate on training by allowing a model to work with data held in their separate controlled environments. DEPA's permissions, safeguards, and provenance rules help govern that collaboration without moving the underlying data.

The article presents responsible access as a major challenge, not as a claim that more data always guarantees a better model. It says privacy, confidentiality, intellectual property, and institutional control concerns often limit sharing. DEPA seeks to make useful collaboration more secure and accountable.