Create sub-class 'StructuralAttack' for traditional (and other) risk measures of target modules
jim-smith opened this issue · comments
Calling it 'StructuralAttacks' as a place holder as this is based on analysing the target model and the train/test set with no need to run any other attacks, for example
- [] residual degrees of freedom
- [] k-anonymity
- [] class disclosure for leaf in tree
- [] equivalent by extension for forests
- Simon;s suggestion about AUC etc. on sacro-ml as issue 13 here
- Kolmogorov-Smirnoff test on confidence per class between training and test set
- runs test??
Include the rule sets from Richard analysis.