AutoML Models for Wireless Signals Classification and their effectiveness against Adversarial Attacks
Classification of wireless signals using AutoML models and understanding their effectiveness towards adversarial attacks.
- AutoML is used to generate models for the classification of wireless signals. AutoKeras library is used to create AutoML deep learning models.
- How to choose a neural network architecture? – A modulation classification example
- Modulation Classification with Deep Learning
- Convolutional Radio Modulation Recognition Networks
- Fast Deep Learning for Automatic Modulation Classification
- Improvements to Modulation Classification Techniques using Deep Learning
- Black-box Adversarial ML Attack on Modulation Classification
Dataset: RML2016.10a.tar.bz2 Source of Dataset: https://www.deepsig.ai/datasets
- All modulation schemes and SNRs of the RadioML dataset are considered for training and testing.
- RadioML dataset is split into training and validation datasets with 20% of data for validation.
- Both training and validation dataset contains samples from all modulation schemes for all SNRs.
- AutoML Customised ResNet, AutoML Customised CLDNN, AutoML Customised CNN and AutoML Customised RNN.
- Models are trained and evaluated on training and validation datasets respectively.
- Performance is also validated on ResNet, CLDNN and Robust-CNN models from Improvements to Modulation Classification Techniques using Deep Learning.
Transfer-Based Untargeted Adversarial Attacks are performed on AutoML Models.
- Projected Gradient Descent(PGD) technique is used to generate adversarial samples using a surrogate model Robust-CNN from Improvements to Modulation Classification Techniques using Deep Learning.
- Carlini and Wagner(CW) technique is used to generate adversarial samples using a surrogate model Robust-CNN from Improvements to Modulation Classification Techniques using Deep Learning.