Privacy Centric AI

I have been working on privacy-centric AI for the past 3 years in different capacities, most recently as CTO at FlintX and for a very long time as a researcher at the Data, Systems and HPC (DaSH) lab at BITS. My work in this area includes both technical and business exploration in depth. My research work in the area got published in a Core A conference, and my startup has received the support of the Hong Kong Science and Technology Park to work in this domain through a grant. I have worked on figuring out the needs of privacy-centric AI model training technologies in the market and have contributed to publications to strengthen privacy-centric research.

Over the past 8 to 10 months at FlintX, I’ve been deeply involved in market research and user validation across a range of ideas, including the use of Federated Learning (FL) in virtual nurse support for cancer patients and its potential applications in the autonomous vehicle industry.

While the technology showed promise, our exploration revealed that the market wasn’t yet favorable for offering Federated Learning as a standalone product. This led us to pivot and explore a new direction: recursive AI, or AI that builds AI.

We envisioned a multi-agent AI framework that could be deployed within an organization’s infrastructure, capable of generating custom AI models based on local needs, while ensuring that sensitive data remains securely on-premise.

UnifyFL is a decentralized framework designed to enable collaboration between organizations using Federated Learning (FL), without depending on a centralized aggregator.

In typical FL setups, machine learning models are trained on private data across different devices or organizations and then sent to a central node (aggregator) for combination. This keeps the data private, but often prevents collaboration between organizations due to trust issues or resource limitations.

UnifyFL addresses this problem by offering a new approach that:

• Eliminates the need for a trusted third-party aggregator.

• Uses decentralized orchestration and distributed storage to ensure secure collaboration.

• Supports both synchronous and asynchronous training modes to handle slow or delayed participants (stragglers).

With this design, UnifyFL empowers organizations to collaboratively build models while maintaining control over their data and optimizing resource usage.

Our evaluations show that UnifyFL delivers performance comparable to centralized FL systems, but with greater trust and flexibility.

• EkatraFL: Enabling Decentralized Cross-Silo Federated Learning | IEEE International Conference on High Performance Computing, Data, and Analytics, 2023