Malware Classifier Backdoors
A study of what happens to a malware classifier when the training data is poisoned, keeping the clean detection rate beside the attack success rate.
Academic research and report. Using and extending an existing backdoor-study implementation.
A trigger that flips a prediction is only half the story. If ordinary detection collapses at the same time then the attack has done something different, so both numbers sit next to each other in the plots.
The manuscript covers several model choices and poisoning conditions, with the feature and backdoor comparisons in the supporting material.
State
A manuscript, figures and comparisons. It's a record of what was measured, and it leaves the general claims alone.