Skip to content

LIMUNIMI/thesis_adversarial_ml_audio

Repository files navigation

A systematic evaluation of adversarial attacks against speech emotion recognition models

Source code for the paper "A systematic evaluation of adversarial attacks against speech emotion recognition models" [Submitted]

Set up

Use conda and the environment file provided

Description of notebooks

  • attacks.ipynb: Run the attacks and the evaluation
  • correleations_analysis.ipynb: Perform preliminary analysis on the datasets for identifying most correlated classes
  • datasets_preprocessing_phase1.ipynb: Perform pre-processing (step 1)
  • datasets_preprocessing_phase2.ipynb: Perform pre-processing (step 2), including data augmentation
  • datasets_sr_test.ipynb:
  • data_visualization.ipynb:
  • defenses-Copy1.ipynb:
  • defenses.ipynb:
  • hyperparameter_optimization-Copy1.ipynb:
  • hyperparameter_optimization.ipynb: Perform the hyper-parameter fine-tuning
  • image_specotrograms.ipynb: Create images and audio of original and attacked samples
  • metadata_preparation.ipynb:
  • model_CNN_LSTM.ipynb: Create the models
  • preprocessing_comparison.ipynb:
  • random_guess_dummy_classifier.ipynb: Run a random guessing classifier on the data

Demo

Here are some audio as examples of the original and attacked samples.

EmoDB

EMOVO

Ravdess

Cite us

[TODO]

Releases

No releases published

Packages

No packages published