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 prevalence and success of Deep Neural Network (DNN) applications inrecent years have motivated research on DNN compression, such as pruning andquantization. These techniques accelerate model inference, reduce powerconsumption, and reduce the size and complexity of the hardware necessary torun DNNs, all with little to no loss in accuracy. However, since DNNs arevulnerable to adversarial inputs, it is important to consider the relationshipbetween compression and adversarial robustness. In this work, we investigatethe adversarial robustness of models produced by several irregular pruningschemes and by 8-bit quantization. Additionally, while conventional pruningremoves the least important parameters in a DNN, we investigate the effect ofan unconventional pruning method: removing the most important model parametersbased on the gradient on adversarial inputs. We call this method GreedyAdversarial Pruning (GAP) and we find that this pruning method results inmodels that are resistant to transfer attacks from their uncompressedcounterparts.

Hardening DNNs against Transfer Attacks during Network Compression using Greedy Adversarial Pruning