"A New Way to Look at Data Privacy"
Researchers at MIT developed a new data privacy metric, Probably Approximately Correct (PAC) Privacy, and an algorithm based on this metric that can automatically determine the minimum amount of randomness that must be added to a Machine Learning (ML) model in order to protect sensitive data such as lung scan images from adversaries. The algorithm does not require knowledge of a model's inner workings or its training process, making it simpler to apply to various models and applications. In multiple cases, the researchers demonstrate that the amount of noise needed to protect sensitive data is significantly less with PAC Privacy than with other approaches. This could help engineers develop ML models that provably hide training data while maintaining accuracy in the real world. This article continues to discuss the privacy technique created by MIT researchers that protects sensitive data while maintaining an ML model's performance.