Practical Efficacy of Zero-Knowledge Verification in Machine Learning
Click the Poster to View Full Screen, Right click to save image
Matthew Fernandez
College:
The Dorothy and George Hennings College of Science, Mathematics, and Technology
Major:
Computer Information Systems
Faculty Research Advisor(s):
Jing-chiou Liou
Abstract:
Though the need for experienced machine learning professionals training and testing AI models will only grow in the coming years, not everyone can afford access to the proper hardware needed for such computationally expensive processes. Outsourcing via cloud computing is an obvious remedy, but it can be vulnerable to tampering from third parties with both the training inputs for these models and the parameters with which they are designed to function. Zero knowledge proofs have been proposed as a means of ensuring this integrity. The novelty of zero knowledge is its ability to prove that the model has been trained with the intended inputs and parameters without revealing the sensitive and often proprietary inputs and parameters themselves. Multiple zero knowledge frameworks in this context have proven to be technically viable, but the growing need for this outsourced machine learning necessitates that cloud service providers evaluate which of the many available zero knowledge frameworks will perform best at scale, while adding minimal cost to the initial investment. We will be comparing technically sound zero knowledge frameworks that have been used to validate the training integrity of the same image dataset, and from this suggest which of these existing systems will have the best performance (speed and memory) for enterprise-level cloud-based zero-knowledge machine learning.