a query about binary search local

[MangoKiller]

Hello,
I am working through your implementation of the OPT attack,
If my understand is not wrong , the part of binary_search_local may waste some opportunities of query：

      if model.predict_label(x0+lbd*theta) == y0:
        lbd_lo = lbd
        lbd_hi = lbd*1.01
        nquery += 1
        while model.predict_label(x0+lbd_hi*theta) == y0:
            lbd_hi = lbd_hi*1.01
            nquery += 1
            if lbd_hi > 20:
                return float('inf'), nquery

In my opinion, if add lbd_lo = lbd_hi , you know, just like :

   if model.predict_label(x0+lbd*theta) == y0:
        lbd_lo = lbd
        lbd_hi = lbd*1.01
        nquery += 1
        while model.predict_label(x0+lbd_hi*theta) == y0:
            lbd_lo = lbd_hi 
            lbd_hi = lbd_hi*1.01
            nquery += 1
            if lbd_hi > 20:
                return float('inf'), nquery

Some opportunities of querying may be saved in the process of binary search.
I wonder if u could clarify this issue for me,
Thanks a miiiiiiiiiiiiillion!

[MangoKiller]

Also, I don't find Normalize([0.485....]) in the process of using CIFAR10, if u could clarify this issue for me, I will very appreciate!

[cmhcbb]

Hi,
Good suggestion. Yes, in terms of finding better lbd_hi, it does find a better initialization. However, I believe the gain would be really marginal.
For the normalization, you could do the normalization for the dataset mean and variance like [0.485...] for a slightly better performance. However, it could be just normalize to image scale to [0,1]. It is easier for us to do the adversarial attack since we don't have clip each dimension separably.

[MangoKiller]

Oh I see, thanks for your help~^_^

[MangoKiller]

Excuse me, another small question is:
Why you did a lot of record on number of query in [OPT_attack, etc] just like:

for i in range(num_directions):
         query_count += 1    #<----------just like this
         theta = np.random.randn(*x0.shape)

I mean, you could use PytorchModel.get_num_queries( )
Thanks~

[cmhcbb]

Yes, you could. It is the legacy code for recording the query_count. A better way would be just call get_num_queries().

[MangoKiller]

I see，again thanks for clarifying