Privately assessing data sharing for machine learning
Original title: Should We Combine Our Data? Algorithms and Limits for Privately Assessing Data Sharing
Interactive communication protocol allows two parties to privately test whether their datasets are compatible for joint training under epsilon-differential privacy using Theta(log log m over epsilon) samples for…
Alice and Bob each hold sensitive labeled data and want to know if pooling increases the number of mistakes opt_H made by any hypothesis in class H on combined data. They distinguish opt_H equals 0 from opt_H greater than or equal to t under pure epsilon-two-party differential privacy without revealing data points. For thresholds over domain of size m the optimal t is tilde Theta(log log m over epsilon). For halfspaces in R to the d the dependence on m becomes log m with protocol t equals tilde O of d cubed log m over epsilon.
- distinguish opt_H equals 0 from opt_H greater than or equal to t
- pure epsilon-two-party differential privacy
- optimal t tilde Theta(log log m over epsilon) for thresholds
- protocol t tilde O(d cubed log m over epsilon) for halfspaces
- some domains require t Omega(d log(m over d) over epsilon)
From the original article
1 min read at arxiv.orgIn this paper, we study when two parties should pool their data to train a machine learning model. Two agents, Alice and Bob, each hold a sensitive labeled dataset and want to train a classifier from a given hypothesis class $\mathcal{H}$.
By Mahdi Haghifam · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.