Mean Observed Dissimilarity

• is the mean of the NISSIM dissimilarity over the adversarial test set for similar levels of attack.
• So for every adversarial set X ∗, calculate NISSIM value for all samples in that set, and divide by the total number of samples.
• (0,1], such that 0 indicates total similarity while 1 indicates total dissimilarity
• $MO{D}_{advset}=\frac{1}{N}\Sigma (NISSI{M}_i)$