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

The huge supporting training data on the Internet has been a key factor inthe success of deep learning models. However, this abundance ofpublic-available data also raises concerns about the unauthorized exploitationof datasets for commercial purposes, which is forbidden by dataset licenses. Inthis paper, we propose a backdoor-based watermarking approach that serves as ageneral framework for safeguarding public-available data. By inserting a smallnumber of watermarking samples into the dataset, our approach enables thelearning model to implicitly learn a secret function set by defenders. Thishidden function can then be used as a watermark to track down third-partymodels that use the dataset illegally. Unfortunately, existing backdoorinsertion methods often entail adding arbitrary and mislabeled data to thetraining set, leading to a significant drop in performance and easy detectionby anomaly detection algorithms. To overcome this challenge, we introduce aclean-label backdoor watermarking framework that uses imperceptibleperturbations to replace mislabeled samples. As a result, the watermarkingsamples remain consistent with the original labels, making them difficult todetect. Our experiments on text, image, and audio datasets demonstrate that theproposed framework effectively safeguards datasets with minimal impact onoriginal task performance. We also show that adding just 1% of watermarkingsamples can inject a traceable watermarking function and that our watermarkingsamples are stealthy and look benign upon visual inspection.

Did You Train on My Dataset? Towards Public Dataset Protection with Clean-Label Backdoor Watermarking