13 12th February, Monday

I've got the capacity attack working on BlueCrystal, but I've found that it takes ages for the attack to learn the malicious features. I'm fairly certain that I could apply some error correction to get the accuracy up, the network seems to get about 30% accuracy quickly, but then any more seems hard to achieve without a large number of epochs, although it does creep up slowly. If I do a kind of data augmentation, to map lots of similar random images to the same output then it might make it a lot easier for the network to learn the malicious features. I think I'd have to get something like this working later on into the project if I want this to work in practice.

I've also started some more writeup in my thesis, although I'm yet to reach two pages yet. This week I'm going to look to get some more writing done but a lot of it will be based on the experiments that I'll be doing.

The BC reservation COSC004081 has been approved, but I still can't seem to get access using it - "srun: error: Unable to allocate resources: Invalid account or account/partition combination specified" is the error BC gives me if I try using it.

Next stages of the project will be to get some secure models working and get some comparison going early to make sure that I'll be getting results that I will be able to talk about.

Update:

I now have the differential privacy code running in tensor flow, turned out to be much easier than expected.

I have continued to write more in my thesis about some of the ideas I am exploring. I'll need to share this some time in the week so that what I'm writing can be reviewed, hopefully I'm starting to become a bit clearer in what I am aiming to do/achieve with this project.