Adversarial robustness evaluation of representation learning models and universal audio representations
Source code for the paper "Adversarial Robustness Evaluation of Representation Learning for Audio Classification".
Use conda and the environment files provided as specified:
- hearPipeline: for the Model_import notebook.
- attackPipeline: for the Main_Loop notebook.
- Dataset_import: Download, import and decompress the HEAR tasks and datasets.
- Resample: Resample the audio files to target sampling rates.
- Model_import: Import the HEAR models, compute and evaluate the embeddings.
- Main_Loop: Perform the attacks, evaluate and present the results.
- SVM: Perform the SVM-based detection of adversarial examples and present the results.
The results are presentend in the notebooks.
For a direct access the two zip files contain the final results for the Attack and SVM phases.