Do we have to assume white box or black box access to the models for the detection of trojans? Like, could we use the gradients and weights of the model in which we have to detect the trojan?
Posted by: mrsarthakgupta @ July 27, 2022, 3:30 p.m.Hello,
Sorry for the late reply. You have white-box access to models at training and test time, so you can use the gradients and weights of the models to help with detection.
All the best,
Mantas (TDC co-organizer)