1 paper · 1 filter
Ethan Wisdom, Tejas Gokhale, Chaowei Xiao +1
In this work, we present a data poisoning attack that confounds machine learning models without any manipulation of the image or label. This is achieved by simply leveraging the mo…