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

Adversarial sample attacks perturb benign inputs to induce DNN misbehaviors.Recent research has demonstrated the widespread presence and the devastatingconsequences of such attacks. Existing defense techniques either assume priorknowledge of specific attacks or may not work well on complex models due totheir underlying assumptions. We argue that adversarial sample attacks aredeeply entangled with interpretability of DNN models: while classificationresults on benign inputs can be reasoned based on the human perceptiblefeatures/attributes, results on adversarial samples can hardly be explained.Therefore, we propose a novel adversarial sample detection technique for facerecognition models, based on interpretability. It features a novelbi-directional correspondence inference between attributes and internal neuronsto identify neurons critical for individual attributes. The activation valuesof critical neurons are enhanced to amplify the reasoning part of thecomputation and the values of other neurons are weakened to suppress theuninterpretable part. The classification results after such transformation arecompared with those of the original model to detect adversaries. Results showthat our technique can achieve 94% detection accuracy for 7 different kinds ofattacks with 9.91% false positives on benign inputs. In contrast, astate-of-the-art feature squeezing technique can only achieve 55% accuracy with23.3% false positives.

Attacks Meet Interpretability: Attribute-steered Detection of Adversarial Samples