-
Notifications
You must be signed in to change notification settings - Fork 130
Tutorials
moo_hax edited this page May 3, 2021
·
8 revisions
Counterfit works to keep the target in focus for the user and tries to provide a uniform interface from which to use the underlying frameworks. However, understanding how to build a class is important for successful use. For a warmup, we will build a target for everyone’s favorite ML model, MNIST.
- Start Counterfit and execute the new command. Enter a name and select images as the data type.
[[email protected]] -> python .\counterfit.py
---------------------------------------------------
Microsoft
__ _____ __
_________ __ ______ / /____ _____/ __(_) /_
/ ___/ __ \/ / / / __ \/ __/ _ \/ ___/ /_/ / __/
/ /__/ /_/ / /_/ / / / / /_/ __/ / / __/ / /
\___/\____/\__,_/_/ /_/\__/\___/_/ /_/ /_/\__/
#ATML
---------------------------------------------------
[+] 18 attacks
[+] 4 targets
counterfit> new
? Target name: mnist
? Which framework? art
? What data type? image
counterfit>
- Find the new target folder in
counterfit/targets, and open the new target python file in your preferred code editor. The code file is generated from a template incounterfit/core/commands/new.py.
# Generated by counterfit #
from counterfit.core.targets import ArtTarget
class Mnist(ArtTarget):
model_name = "mnist"
model_data_type = "image"
model_endpoint = ""
model_input_shape = ()
model_output_classes = []
X = []
def __init__(self):
self.X = []
def __call__(self, x):
return x- In your code editor, fill out the required target properties.
-
model_nameandmodel_data_typewere taken care of during new target creation. -
model_endpointis where Counterfit will collect outputs from the target model. We will use themnist_sklearn_pipeline.pklpre-trained model found in the tutorial folder. -
model_input_shapeis the input shape of the target model, which is a known(1, 28, 28). -
model_output_classesare the output classes of the model, which is a known["0", "1", "2", "3", "4", "5", "6", "7", "8", "9"].
After filling in the blanks with the information above, the target class should look like the following,
# Generated by counterfit #
from counterfit.core.targets import ArtTarget
class Mnist(ArtTarget):
model_name = "mnist"
model_data_type = "image"
model_endpoint = "counterfit/targets/tutorial/mnist_sklearn_pipeline.pkl"
model_input_shape = (1, 28, 28)
model_output_classes = ["0", "1", "2", "3", "4", "5", "6", "7", "8", "9"]
X = []
def __init__(self):
self.X = []
def __call__(self, x):
return x- Interact with the target via
interact. Try andreloadthe target, fix any errors that show up. Now when youlist targetsome of the information should be filled out.
counterfit> interact mnist
...
mnist> reload
mnist> list targets
Name Type Input Shape Location
----------------------------------------------------------------------------------------------------------------------------------------
creditfraud numpy (30,) counterfit/targets/creditfraud/creditfraud_sklearn_pipeline.pkl
mnist image (1, 28, 28) counterfit/targets/tutorial/mnist_sklearn_pipeline.pkl
moviereviews text (1,) counterfit/targets/moviereviews/movie_reviews_sentiment_analysis.pt
satelliteimages image (3, 256, 256) counterfit/targets/satelliteimages/satellite-image-params-airplane-stadium.h5
tutorial image (1, 28, 28) counterfit/targets/tutorial/mnist_sklearn_pipeline.pkl
mnist>
- With the required properties in place, we can start loading resources and implementing functionality.
- This model is an
imagedata type. A user can overrideclip_valuesin the target. This ensures image values remain valid pixel values. - Because this is a local model, we first load the model and expose the
predictfunction that Counterfit will use to interact with the target model. - Next, load sample data
X. The sample data is a list of lists where each list is an array containing a processed sample. The data for the tutorial is in a nice tidy numpy zip file, however, most targets will require additional processing to getX.
Paste the __init__ function below in the target class.
def __init__(self):
self.clip_values = (0, 255)
with open(self.model_endpoint, "rb") as f:
self.model = pickle.load(f)
self.data_file = "counterfit/targets/tutorial/mnist_784.npz"
self.sample_data = np.load(self.data_file, allow_pickle=True)
self.X = self.sample_data["X"]- Excellent, we now have samples and a model to attack. Next, we will build the
__call__function, Counterfit will use this function to submit samples to the target model viax.xis a perturbed sample of shape(Batch, Channels, Height, Width). Channels, Height, and Width are derived from themodel_input_shapethat was defined earlier. Functionally,xis a list of lists, where each list is a sample of shape(1, 28, 28). This should sound familiar as it is the same shape asXandmodel_input_shape. Paste the following code below the__init__function.
def __call__(self, x):
scores = self.model.predict_proba(x.reshape(x.shape[0], -1))
return scores.tolist()Note: A crucial piece to the __call__ function is how scores are returned to the attack algorithm. The must return a list of probabilities. An attack algorithm uses the returned scores to inform how to change the sample for the next iteration of the attack. In this tutorial the pre-trained MNIST model returns exactly what is needed, which is a list of probabilities for each label.
- Alright, the new target is almost ready. Add the following imports to the top on the file,
import pickle,import numpy as np. Next, execute thereloadcommand to load the updated target into the session. The__init__function is called onreloadorinteract. There will be a warnings that are not suppressed to keep the target code clean - you can safely ignore them. , The final target should look like below,
# Generated by counterfit #
import pickle
import numpy as np
from counterfit.core.targets import ArtTarget
class Mnist(ArtTarget):
model_name = "mnist"
model_data_type = "image"
model_endpoint = "counterfit/targets/tutorial/mnist_sklearn_pipeline.pkl"
model_input_shape = (1, 28, 28)
model_output_classes = ["0", "1", "2", "3", "4", "5", "6", "7", "8", "9"]
X = []
def __init__(self):
self.clip_values = (0, 255)
with open(self.model_endpoint, "rb") as f:
self.model = pickle.load(f)
self.data_file = "counterfit/targets/tutorial/mnist_784.npz"
self.sample_data = np.load(self.data_file, allow_pickle=True)
self.X = self.sample_data["X"]
def __call__(self, x):
scores = self.model.predict_proba(x.reshape(x.shape[0], -1))
return scores.tolist()- To test the functionality of the target, execute the
predictfunction.
mnist> predict
[!] No index sample, setting random index.
Output Scores
Sample ['0' '1' '2' '3' '4' '5' '6'
Index Sample '7' '8' '9']
------------------------------------------------------------------------------------------------------
65923 mnist-sample-46600446.png [0.000 0.000 0.000 0.000 1.000
0.000 0.000 0.000 0.000 0.000]
mnist>
- Excellent. We are ready to run attacks on the MNIST target. List the frameworks, load art, and then list the available attacks. Attacks are filtered based on the
model_data_typedefined in the target class.
counterfit> list frameworks
Framework # of Attacks
----------------------------------------------------
art 7
textattack 11
counterfit> load art
[+] Framework loaded successfully!
counterfit> list attacks
Name Type Category Tags Framework
----------------------------------------------------------------------------------------
boundary evasion blackbox image, numpy art
hop_skip_jump evasion blackbox image, numpy art
pixel evasion blackbox image art
spatial_transformation evasion blackbox image art
square evasion blackbox image art
threshold evasion blackbox image art
zoo evasion blackbox image, numpy art
counterfit>
- Add an attack to the pipeline by executing
use hop_skip_jump.
mnist> use hop_skip_jump
[+] Using hop_skip_jump c596a8f3
mnist>hop_skip_jump>
- Finally, start the attack with
run.
mnist>hop_skip_jump> run
[+] Running hop_skip_jump on mnist
HopSkipJump: 100%|█████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:32<00:00, 32.74s/it]
[+] 1/1 succeeded
Sample Index Label (conf) Attack Label (conf) % Eucl. dist. Elapsed Time [sec] Queries (rate) Attack Input
---------------------------------------------------------------------------------------------------------------------------------
1. 0 5 (0.9990) 3 (0.6320) 0.02039% 32.8 24548 (749.4 counterfit/ta
query/sec) rgets/mnist/r
esults/mnist-
c596a8f3-fina
l-0-label-3.p
ng
mnist>hop_skip_jump>
- Alternatively, run multiple attacks with
scan. Issue thebackcommand to exit the active attack. Then usescan,
mnist>hop_skip_jump> scan --iterations 2 --attack hop_skip_jump
[+] Running these attacks 2x each:
hop_skip_jump
[+] Using hop_skip_jump f36ede50
HopSkipJump: 100%|█████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:04<00:00, 4.07s/it]
[+] Using hop_skip_jump 196a6995
HopSkipJump: 100%|█████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 16.58it/s]
===============
SCAN SUMMARY
===============
Time[sec] Queries Best Score
Attack Name Total Runs Successes (%) (min/avg/max) (min/avg/max) (attack_id) Best Parameters
---------------------------------------------------------------------------------------------------------------------------
hop_skip_jump 2 1 (50.0%) 0.1/ 2.1/ 4.1 51/ 1746/ 3441 1.0 (f36ede50) init_eval=78
init_size=29
max_eval=3631
max_iter=15
norm=inf
targeted=false
sample_index=23298
target_class=6
mnist>hop_skip_jump>
- Save the results with
save.
mnist>hop_skip_jump> save
[+] Successfully wrote counterfit/targets/mnist/results/mnist_9f806eec.json
mnist>hop_skip_jump>