![]() ![]() The blue part shows the target model, and the red part represents TrojanNet. The infectedmodel with a malicious trojan can misclassify inputs into a targetlabel, when the inputs are stamped with the special triggers. Specifically,we do not change parameters in the original model but insert atiny trojan module (TrojanNet) into the target model. Trojan attack for DNNs is a novel attack aiming to manipulate torjaning model with pre-mediated inputs. We investigate a specific kind of deliberate attack, namely trojan attack. Livestation viewers can add channels from a directory, and may add their own choice of channels to the directory (limited to Windows Media streams).This is the keras implemention for KDD2020 paper “An Embarrassingly Simple Approach for Trojan Attack in Deep Neural Networks” this paper ( bibtex here for citation). Despite advertisement boasting Deutsche Welle TV, this channel is not available in some locations, including but not limited to the United States. ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |