19 22nd February, Thursday

Some interesting discoveries! It seems that having applied the capacity abuse attack to the differentially private networks I've needed to use much stronger privacy than expected to limit the attack. In fact, the delta value of the privacy also seems to be much more important and the amortized approach that appears in the open source code but not in the paper also seems to be more effective than the main method they outline.

I will continue to investigate this to get a better feel for the effects of tuning different parameters and will try to get some hard results to look at.