English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Within the realm of privacy-preserving machine learning, empirical privacydefenses have been proposed as a solution to achieve satisfactory levels oftraining data privacy without a significant drop in model utility. Mostexisting defenses against membership inference attacks assume access toreference data, defined as an additional dataset coming from the same (or asimilar) underlying distribution as training data. Despite the common use ofreference data, previous works are notably reticent about defining andevaluating reference data privacy. As gains in model utility and/or trainingdata privacy may come at the expense of reference data privacy, it is essentialthat all three aspects are duly considered. In this paper, we first examine theavailability of reference data and its privacy treatment in previous works anddemonstrate its necessity for fairly comparing defenses. Second, we propose abaseline defense that enables the utility-privacy tradeoff with respect to bothtraining and reference data to be easily understood. Our method is formulatedas an empirical risk minimization with a constraint on the generalizationerror, which, in practice, can be evaluated as a weighted empirical riskminimization (WERM) over the training and reference datasets. Although weconceived of WERM as a simple baseline, our experiments show that,surprisingly, it outperforms the most well-studied and current state-of-the-artempirical privacy defenses using reference data for nearly all relative privacylevels of reference and training data. Our investigation also reveals thatthese existing methods are unable to effectively trade off reference dataprivacy for model utility and/or training data privacy. Overall, our workhighlights the need for a proper evaluation of the triad model utility /training data privacy / reference data privacy when comparing privacy defenses.

A Cautionary Tale: On the Role of Reference Data in Empirical Privacy Defenses