naiyeleo / abs Goto Github PK

Any plan to make the keras version open-source?

I am trying to compare the performance gap between the Nerual Cleanse and the ABS under the feature space attack. It looks like all .pkl files in ./deltas are for pixel-type triggers. How to set a proper trigger for the filter_stamp?

Recently I used the pytorch version for some simple backdoor attacks, and it really works well.
However, I have a question when reading your paper and the code:
In sec 4.2 and previous sections, the paper claims that 'the compromised neurons should activate in a specific range' to get the output elevation. I think this statement makes sense and is intuitive. However, in algorithm 2 and the code, it seems that you maximum the activation value of the compromised neuron, instead of letting them activate in the specific range, in which could gets the required output elevation.
So, is maximizing the activation value of the compromised neurons equivalent to activating them within a specific range? Is it because I missed some analysis and misunderstood it?

Dear Author,
I am trying to reproduce your results in my experiments.

However, I am unable to use "python2 run_abs.py". The error I get is given as follows with a full traceback:

Traceback (most recent call last):
File "run_abs.py", line 1, in
import abs
File "abs.pyx", line 29, in init abs
AttributeError: 'module' object has no attribute 'set_random_seed'

I have tried downgrading my tensorflow to 1.12.0, but it doesn't help. Would appreciate your help here.

Thanks

Hi,

Can you please add the dependencies for the TrojAI versions?

Also, if I want to experiment with your method on another dataset, which version should I use? I would guess TrojAI round4 to be the one since it is the latest. But, your confirmation/correction is highly appreciated.

Best,

I was trying to run the round 4 code at TrojAI_competition/round4/abs_pytorch1_color_simple_filter4_11_9_d_classifier.py

There is a directory defined as a global variable: trained_triggers_dir = '/trained_triggers'
Seems files in that folder are used during filter check, for example: delta_name = trained_triggers_dir+'/gotham/deltas/id-00000142_model_18_464_900_4_11.pkl'

I got some errors while running as the directory ''/trained_triggers'' doesn't exist. I didn't find another directory with those filter files either. Would you mind uploading the files? or should I generate the files in some way? Thank you :)