More New AI Backdoor Papers!

[taylor.kul...@iarpa.gov]

Hello TrojAI Community, 

Some new papers were released to arXiv recently that I thought were interesting and wanted to share! 

Poisoned Classifier are not only backdoored, they are fundamentally broken

One Sentence Summary: Outlines a method for finding alternative triggers in poisoned models and proposes that backdoor attacks make neural networks susceptible to a wide range of backdoors in addition to the one directly inserted by the adversary 

Input-Aware Dynamic Backdoor Attack

One Sentence Summary:  Proposes a new attack where the trigger varies from input to input in an attempt to make backdoor verification more difficult 

Feel free to any other papers you have found!

Best,

Taylor Kulp-McDowall

Tech SETA, TrojAI

taylor.kul...@iarpa.gov