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

Existing data-free model stealing methods use a generator to produce samplesin order to train a student model to match the target model outputs. To thisend, the two main challenges are estimating gradients of the target modelwithout access to its parameters, and generating a diverse set of trainingsamples that thoroughly explores the input space. We propose a Dual Studentmethod where two students are symmetrically trained in order to provide thegenerator a criterion to generate samples that the two students disagree on. Onone hand, disagreement on a sample implies at least one student has classifiedthe sample incorrectly when compared to the target model. This incentivetowards disagreement implicitly encourages the generator to explore morediverse regions of the input space. On the other hand, our method utilizesgradients of student models to indirectly estimate gradients of the targetmodel. We show that this novel training objective for the generator network isequivalent to optimizing a lower bound on the generator's loss if we had accessto the target model gradients. We show that our new optimization frameworkprovides more accurate gradient estimation of the target model and betteraccuracies on benchmark classification datasets. Additionally, our approachbalances improved query efficiency with training computation cost. Finally, wedemonstrate that our method serves as a better proxy model for transfer-basedadversarial attacks than existing data-free model stealing methods.

Dual Student Networks for Data-Free Model Stealing