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

Quantization is a popular technique that $transforms$ the parameterrepresentation of a neural network from floating-point numbers intolower-precision ones ($e.g.$, 8-bit integers). It reduces the memory footprintand the computational cost at inference, facilitating the deployment ofresource-hungry models. However, the parameter perturbations caused by thistransformation result in $behavioral$ $disparities$ between the model beforeand after quantization. For example, a quantized model can misclassify sometest-time samples that are otherwise classified correctly. It is not knownwhether such differences lead to a new security vulnerability. We hypothesizethat an adversary may control this disparity to introduce specific behaviorsthat activate upon quantization. To study this hypothesis, we weaponizequantization-aware training and propose a new training framework to implementadversarial quantization outcomes. Following this framework, we present threeattacks we carry out with quantization: (i) an indiscriminate attack forsignificant accuracy loss; (ii) a targeted attack against specific samples; and(iii) a backdoor attack for controlling the model with an input trigger. Wefurther show that a single compromised model defeats multiple quantizationschemes, including robust quantization techniques. Moreover, in a federatedlearning scenario, we demonstrate that a set of malicious participants whoconspire can inject our quantization-activated backdoor. Lastly, we discusspotential counter-measures and show that only re-training consistently removesthe attack artifacts. Our code is available athttps://github.com/Secure-AI-Systems-Group/Qu-ANTI-zation

Qu-ANTI-zation: Exploiting Quantization Artifacts for Achieving Adversarial Outcomes