Sparse Approximations for High Fidelity Compression of Network Usenix Form

Instructions and help about Sparse Approximations For High Fidelity Compression Of Network Usenix

we'll be talking about our paper high accuracy and high fidelity extraction of neural networks this is joint work with nicolas carlini david brechtlow alex karakin and nicola paperno this work was done while all of us were at google research but i've since moved back to my home institution of northeastern university and nicolas has moved to university of toronto our talk starts with a common setup machine learning as a service digit co has trained a very good handwritten digit recognition model and alice is willing to pay to use it alice identifies some handwritten five sends it to digit co digico runs it through the model finds out it's a five and sends that information back to alice however if alice is malicious she can use this input output interaction to steal the model this attack is known as model extraction and is the focus of our talk today i'm going to be discussing three parts of our paper a taxonomy we introduce for model extraction and improvements to two types of extractio