02 29th January, Monday

I can now plot nice histograms for the weight distributions in all layers of the smallest instance of resnet in Lasagne, as implemented in the ml-models code. I tested the network with no attack for 8 epochs and then the correlation encoding attack for 10 epochs (first model in the morning and second in the afternoon- both reached about 75% validation accuracy). I chose the correlation encoding attack as it appeared to give the best distributions.

Interesting, re-reading the paper today I noticed that it is briefly mentioned that model compression could be a way of dealing with some of these attacks. This could potentially be something to investigate, by rerunning these attacks against a compressed model I would be able to investigate this claim.