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 Tools annual

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 Tools monthly

Global: Zendy Plus monthly

Presents the access of premium content as premium feature

Premium Content

Global: Zendy Plus annual

Global: Zendy Free Plan

It is known that deep neural networks, trained for the classification ofnon-sensitive target attributes, can reveal sensitive attributes of their inputdata through internal representations extracted by the classifier. We take astep forward and show that deep classifiers can be trained to secretly encode asensitive attribute of their input data into the classifier's outputs for thetarget attribute, at inference time. Our proposed attack works even if usershave a full white-box view of the classifier, can keep all internalrepresentations hidden, and only release the classifier's estimations for thetarget attribute. We introduce an information-theoretical formulation for suchattacks and present efficient empirical implementations for traininghonest-but-curious (HBC) classifiers: classifiers that can be accurate inpredicting their target attribute, but can also exploit their outputs tosecretly encode a sensitive attribute. Our work highlights a vulnerability thatcan be exploited by malicious machine learning service providers to attacktheir user's privacy in several seemingly safe scenarios; such as encryptedinferences, computations at the edge, or private knowledge distillation.Experimental results on several attributes in two face-image datasets show thata semi-trusted server can train classifiers that are not only perfectly honestbut also accurately curious. We conclude by showing the difficulties indistinguishing between standard and HBC classifiers, discussing challenges indefending against this vulnerability of deep classifiers, and enumeratingrelated open directions for future studies.

Honest-but-Curious Nets: Sensitive Attributes of Private Inputs Can Be Secretly Coded into the Classifiers' Outputs