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

In recent years, machine learning models have been shown to be vulnerable tobackdoor attacks. Under such attacks, an adversary embeds a stealthy backdoorinto the trained model such that the compromised models will behave normally onclean inputs but will misclassify according to the adversary's control onmaliciously constructed input with a trigger. While these existing attacks arevery effective, the adversary's capability is limited: given an input, theseattacks can only cause the model to misclassify toward a single pre-defined ortarget class. In contrast, this paper exploits a novel backdoor attack with amuch more powerful payload, denoted as Marksman, where the adversary canarbitrarily choose which target class the model will misclassify given anyinput during inference. To achieve this goal, we propose to represent thetrigger function as a class-conditional generative model and to inject thebackdoor in a constrained optimization framework, where the trigger functionlearns to generate an optimal trigger pattern to attack any target class atwill while simultaneously embedding this generative backdoor into the trainedmodel. Given the learned trigger-generation function, during inference, theadversary can specify an arbitrary backdoor attack target class, and anappropriate trigger causing the model to classify toward this target class iscreated accordingly. We show empirically that the proposed framework achieveshigh attack performance while preserving the clean-data performance in severalbenchmark datasets, including MNIST, CIFAR10, GTSRB, and TinyImageNet. Theproposed Marksman backdoor attack can also easily bypass existing backdoordefenses that were originally designed against backdoor attacks with a singletarget class. Our work takes another significant step toward understanding theextensive risks of backdoor attacks in practice.

Marksman Backdoor: Backdoor Attacks with Arbitrary Target Class