<ul><li>Backdoor attacks on text classifiers can be made more effective and subtle by crafting trigger attributes that are indistinguishable from normal texts.</li><li>Previous attacks often rely on triggers that are ungrammatical or unusual, making them easily detectable by human annotators during manual inspection.</li><li>The study proposes 'AttrBkd', a method for crafting subtle trigger attributes by extracting fine-grained attributes from existing backdoor attacks.</li><li>Human evaluations show that AttrBkd with baseline-derived attributes is more effective and subtle compared to original baseline backdoor attacks.</li></ul>

The Ultimate Cookbook for Invisible Poison: Crafting Subtle Clean-Label Text Backdoors with Style Attributes

Discover more