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 attacks have exposed serious vulnerabilities in Deep NeuralNetworks (DNNs) through their ability to force misclassifications throughhuman-imperceptible perturbations to DNN inputs. We explore a new direction inthe field of adversarial attacks by suggesting attacks that aim to degrade thecomputational efficiency of DNNs rather than their classification accuracy.Specifically, we propose and demonstrate sparsity attacks, which adversarialmodify a DNN's inputs so as to reduce sparsity (or the presence of zero values)in its internal activation values. In resource-constrained systems, a widerange of hardware and software techniques have been proposed that exploitsparsity to improve DNN efficiency. The proposed attack increases the executiontime and energy consumption of sparsity-optimized DNN implementations, raisingconcern over their deployment in latency and energy-critical applications.  We propose a systematic methodology to generate adversarial inputs forsparsity attacks by formulating an objective function that quantifies thenetwork's activation sparsity, and minimizing this function using iterativegradient-descent techniques. We launch both white-box and black-box versions ofadversarial sparsity attacks on image recognition DNNs and demonstrate thatthey decrease activation sparsity by up to 1.82x. We also evaluate the impactof the attack on a sparsity-optimized DNN accelerator and demonstratedegradations up to 1.59x in latency, and also study the performance of theattack on a sparsity-optimized general-purpose processor. Finally, we evaluatedefense techniques such as activation thresholding and input quantization anddemonstrate that the proposed attack is able to withstand them, highlightingthe need for further efforts in this new direction within the field ofadversarial machine learning.

Sparsity Turns Adversarial: Energy and Latency Attacks on Deep Neural Networks