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

Despite achieving great success, Deep Neural Networks (DNNs) are vulnerableto adversarial examples. How to accurately evaluate the adversarial robustnessof DNNs is critical for their deployment in real-world applications. An idealindicator of robustness is adversarial risk. Unfortunately, since it involvesmaximizing the 0-1 loss, calculating the true risk is technically intractable.The most common solution for this is to compute an approximate risk byreplacing the 0-1 loss with a surrogate one. Some functions have been used,such as Cross-Entropy (CE) loss and Difference of Logits Ratio (DLR) loss.However, these functions are all manually designed and may not be well suitedfor adversarial robustness evaluation. In this paper, we leverage AutoML totighten the error (gap) between the true and approximate risks. Our maincontributions are as follows. First, AutoLoss-AR, the first method to searchfor surrogate losses for adversarial risk, with an elaborate search space, isproposed. The experimental results on 10 adversarially trained modelsdemonstrate the effectiveness of the proposed method: the risks evaluated usingthe best-discovered losses are 0.2% to 1.6% better than those evaluated usingthe handcrafted baselines. Second, 5 surrogate losses with clean and readableformulas are distilled out and tested on 7 unseen adversarially trained models.These losses outperform the baselines by 0.8% to 2.4%, indicating that they canbe used individually as some kind of new knowledge. Besides, the possiblereasons for the better performance of these losses are explored.

Tightening the Approximation Error of Adversarial Risk with Auto Loss Function Search