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

Bit-flip attacks (BFAs) have attracted substantial attention recently, inwhich an adversary could tamper with a small number of model parameter bits tobreak the integrity of DNNs. To mitigate such threats, a batch of defensemethods are proposed, focusing on the untargeted scenarios. Unfortunately, theyeither require extra trustworthy applications or make models more vulnerable totargeted BFAs. Countermeasures against targeted BFAs, stealthier and morepurposeful by nature, are far from well established.  In this work, we propose Aegis, a novel defense method to mitigate targetedBFAs. The core observation is that existing targeted attacks focus on flippingcritical bits in certain important layers. Thus, we design a dynamic-exitmechanism to attach extra internal classifiers (ICs) to hidden layers. Thismechanism enables input samples to early-exit from different layers, whicheffectively upsets the adversary's attack plans. Moreover, the dynamic-exitmechanism randomly selects ICs for predictions during each inference tosignificantly increase the attack cost for the adaptive attacks where alldefense mechanisms are transparent to the adversary. We further propose arobustness training strategy to adapt ICs to the attack scenarios by simulatingBFAs during the IC training phase, to increase model robustness. Extensiveevaluations over four well-known datasets and two popular DNN structures revealthat Aegis could effectively mitigate different state-of-the-art targetedattacks, reducing attack success rate by 5-10$\times$, significantlyoutperforming existing defense methods.

Aegis: Mitigating Targeted Bit-flip Attacks against Deep Neural Networks