Logging of distortion

[w-garcia]

Hello,

I am working through your implementation of the OPT attack, and had a question about calculation of distortion. It is logged as only g_theta here: https://github.com/cmhcbb/attackbox/blob/master/attack/OPT_attack.py#L69

However if I understand correctly, the right way to log distortion would be as LA.norm(g_theta * best_theta), since the final adversarial sample is x0 + g_theta*best_theta. It is printed that way here: https://github.com/cmhcbb/attackbox/blob/master/attack/OPT_attack.py#L99

Could this be clarified by the authors?

Thanks

[cmhcbb]

Hi,

attackbox/attack/OPT_attack.py

Line 69 in de692af

self.log[0][0], self.log[0][1] = g_theta, query_count

This line is pointing to the first iteration of log, i.e. the distortion after finding the initial points by adding random distortion.
The log records all iteration's distortion and its # of quires until the algorithm stops, so you could think it is a matrix per example as shown in

attackbox/attack/OPT_attack.py

Line 152 in de692af

self.log[i+1:,0] = g_theta

.

Also, the best_theta is always a norm-1 vector so we don't need to calculate the norm again.