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

The backdoor or Trojan attack is a severe threat to deep neural networks(DNNs). Researchers find that DNNs trained on benign data and settings can alsolearn backdoor behaviors, which is known as the natural backdoor. Existingworks on anti-backdoor learning are based on weak observations that thebackdoor and benign behaviors can differentiate during training. An adaptiveattack with slow poisoning can bypass such defenses. Moreover, these methodscannot defend natural backdoors. We found the fundamental differences betweenbackdoor-related neurons and benign neurons: backdoor-related neurons form ahyperplane as the classification surface across input domains of all affectedlabels. By further analyzing the training process and model architectures, wefound that piece-wise linear functions cause this hyperplane surface. In thispaper, we design a novel training method that forces the training to avoidgenerating such hyperplanes and thus remove the injected backdoors. Ourextensive experiments on five datasets against five state-of-the-art attacksand also benign training show that our method can outperform existingstate-of-the-art defenses. On average, the ASR (attack success rate) of themodels trained with NONE is 54.83 times lower than undefended models understandard poisoning backdoor attack and 1.75 times lower under the naturalbackdoor attack. Our code is available athttps://github.com/RU-System-Software-and-Security/NONE.

Training with More Confidence: Mitigating Injected and Natural Backdoors During Training