Create AudioDynamicTrigger #2328
Conversation
This code is designed to use the DynamicTrigger class to create a dynamic backdoor attack by generating a trigger and inserting it into an audio signal. If the generate_dynamic_trigger function correctly returns an instance of the DynamicTrigger class, then the attack works
|
Hi @OrsonTyphanel93 Thank you very much for your pull request! Could you please change the target branch to one of the dev branches? |
Codecov ReportAttention:
❗ Your organization needs to install the Codecov GitHub app to enable full functionality. Additional details and impacted files@@ Coverage Diff @@
## dev_1.18.0 #2328 +/- ##
===============================================
- Coverage 85.62% 70.70% -14.93%
===============================================
Files 324 324
Lines 29323 29331 +8
Branches 5405 5028 -377
===============================================
- Hits 25108 20738 -4370
- Misses 2837 7446 +4609
+ Partials 1378 1147 -231
|
from sklearn.preprocessing import QuantileTransformer
from tensorflow.keras.layers import Input, Dense
from tensorflow.keras.models import Model
def anonymize_speaker(self, spectrogram, noise_std=0.1):
# Create a model for differentially private feature extraction
input_layer = Input(shape=(spectrogram.shape[1],))
hidden_layer = Dense(128, activation='relu')(input_layer)
output_layer = Dense(spectrogram.shape[1])(hidden_layer)
autoencoder = Model(input_layer, output_layer)
autoencoder.compile(optimizer='adam', loss='mean_squared_error')
# Train the autoencoder with noise layers
noisy_spectrogram = spectrogram + np.random.normal(0, noise_std, spectrogram.shape)
autoencoder.fit(noisy_spectrogram, spectrogram, epochs=10, batch_size=32)
# Use the autoencoder to extract features from the spectrogram
features = autoencoder.predict(spectrogram)
# Apply quantization-based transformation
transformer = QuantileTransformer(n_quantiles=100, random_state=0)
quantized_features = transformer.fit_transform(features)
# Reconstruct the spectrogram from the quantized features
reconstructed_spectrogram = autoencoder.predict(quantized_features)
return reconstructed_spectrogram |
|
Hi @beat-buesser please , next time you test codecov, please try DynamicTrigger with this new anonymiser_speaker function, I think it will make code optimization faster. Thanks ! To make the anonymize_speaker method complex, I have incorporated several techniques and concepts such as : **Differentially private feature extraction: We can introduce differentially private feature extractors based on an autoencoder and an automatic speech recognizer, trained using noise layers. This approach was proposed in the paper (Differentially Private Speaker Anonymization see arxiv link ) and has been shown to obtain private utterances with a provable upper bound on the speaker information they contain. ** Quantization-based transformation: We can promote anonymization algorithms based on quantization-based transformation as an alternative to the most widely used and well-known noise-based approach. This approach was proposed in the paper (Anonymizing Speech: Evaluating and Designing Speaker Anonymization Techniques see arxiv link link and can reduce the speaker's PPI (privacy preserving information) while maintaining utility. ** |
|
I recently generalized the DynamicTrigger attack to very complex databases and it succeeded in corrupting all of them without exception. See an example with TIMIT Darpa Dataset speech for speaker verification.
|
|
I'll change the target branch to |
|
Hi guys, I'm doing it, but I don't have access to the 1.18 target! Do you have the possibility to change it directly by yourself? |
Description : Dynamic Stacking of triggers :
https://github.com/OrsonTyphanel93/adversarial-robustness-toolbox/blob/dev_1.14.0/Update_ART_dynamic_backdoor_attacks_trigger_stacking_(_Audio).ipynb
"DynamicTrigger", exploits a "trigger stacking" technique combining numerous triggers to make detection more difficult. The model can learn to correlate the combined trigger with the desired output by stacking triggers. This means that even if the input has only one trigger, the model can anticipate the expected result. As a result, the model has the ability to generate identical samples with similar class titles, or to assign the label designated by the attacker for each sample, depending on its objectives.
This code is designed to use the DynamicTrigger class to create a dynamic backdoor attack by generating a trigger and inserting it into an audio signal. If the generate_dynamic_trigger function correctly returns an instance of the DynamicTrigger class, then the attack works
Test Configuration: