SBIR-STTR Award

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) project includes advances in scientific understanding and substantial societal and commercial impacts. In an era with seemingly endless data breaches, the project offers a way of applying the power of machine learning while never disclosing sensitive raw data. Decentralized computation can increase the scale of models that may be trained, which will allow the use of deep learning on more complicated problems across a range of fields. Additionally, allowing confidential data to be used will allow more rapid research advances in fields with sensitive data, such as biomedicine. Furthermore, decentralized computation offers the promise of lower cost than existing computational infrastructures such as cloud providers. This greater, and more democratic, power will push the boundaries of the state-of-the-art and also enable more people to leverage large-scale machine learning. This SBIR Phase I project proposes to advance knowledge in the area of coordinating decentralized secure machine learning with a blockchain in a manner that maintains data confidentiality and ensures verifiability. The R&D will also advance understanding and practicality of zero knowledge computational verification and homomorphic neural networks. While deep neural networks have yielded astounding results in recent years, there has been limited progress towards achieving a practical solution to training models in a decentralized context while both maintaining data confidentiality and ensuring verifiability. This is the key challenge and it is anticipated that this project will yield a solution. The proposed approach involves defining a protocol for training amongst untrusted parties that is mediated by a decentralized ledger and involves the use of homomorphic encryption and a computational verification technique. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

The broader impact/commercial potential of this Small Business Technology Transfer (STTR) Phase II project will enable collaboration with sensitive data. This will increase the ability of companies and institutions in sensitive domains, including biomedicine, to apply existing machine learning models and increase their ability to train new, more sophisticated models, thereby improving insights for enhanced care. This project will advance a blockchain technique to maintain the integrity of confidential data for shared development of machine learning techniques. This STTR Phase II project proposes to advance knowledge in the area of coordinating decentralized machine learning with a distributed ledger in a manner that maintains data confidentiality and ensures verifiability. This project extends the utility of deep neural networks in domains requiring data privacy, such as partnerships or settings involving decentralized computation. The project focuses on both inference and joint model training. The technical approach will leverage techniques from fully homomorphic encryption (FHE) and secure multiparty computation (MPC). The work will advance an FHE cryptosystem, assess practical performance of candidate cryptosystems, and develop efficient cryptographic verification techniques mediated by the distributed ledger. In addition to creating critical knowledge in the preceding areas, other technical results will include accessible open-source tooling for neural network inference on encrypted data and a modular framework for practical deployment, including interfacing with data stores. The program of work will lead to technology that enables verifiable confidential inference and joint model training and will demonstrate these capabilities on real-world analyses in medicine with medical image and non-image data.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.