The Cost of Differential Privacy in Language Models: The Boundary Between Theory and Practice

Introduction / Main Floor: The Cost of Differential Privacy in Language Models: The Boundary Between Theory and Practice

Recent research reveals the precise cost of implementing differential privacy in language recognition and generation tasks, finding that approximate differential privacy can protect privacy with almost no cost, while the cost of pure differential privacy has also been accurately quantified.