Adversarial Examples for Improving Eye-State Classification

This project studies how adversarial examples can be used to improve the robustness of eye-state classification models. The work compares standard training, adversarial augmentation, and adversarial training strategies, with a focus on models such as Wide Residual Networks and Parseval Networks.

Goals

First objective

Improve model robustness by incorporating adversarial examples into the training data and evaluating the resulting classifier on clean test data.

Second objective

Increase robustness and accuracy against small input perturbations by training with adversarial examples and evaluating performance under attacked test data.

Methodology

• Develop Wide Residual Network and Parseval Network architectures.
• Train neural networks on the eye-state dataset.
• Generate adversarial examples using FGSM and random noise.
• Compare clean-data performance and adversarial robustness.
• Use signal-to-noise ratio (SNR) and transferability as evaluation metrics.

Models

Wide Residual Network

A residual CNN baseline used for comparison.

Parseval Network

A constrained network designed to improve robustness, using:

• Orthogonality constraint in convolutional layers
• Convexity constraint in aggregation layers

Visualization and examples

• Adversarial image notebook
• ROC curve scripts
• Learning curve scripts

Repository layout

adversarial_examples_parseval_net/
├── README.md
├── requirements.txt
├── documents/
│   ├── Expose/
│   ├── Report/
│   └── slide/
├── src/
│   ├── data/
│   ├── logs/
│   ├── models/
│   ├── preprocessing/
│   ├── train/
│   └── visualization/
└── LICENSE

Environment and setup

The project depends on Python scientific, visualization, and deep-learning libraries.

Install dependencies

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install -r requirements.txt

Current environment note

The current workspace was checked with Python 3.13.12, and the pinned package versions in requirements.txt match the installed environment:

• numpy==2.4.6
• pandas==3.0.3
• scikit-learn==1.9.0
• scipy==1.17.1
• matplotlib==3.11.0

TensorFlow and CleverHans may require an environment-specific compatible install depending on the Python version and platform. The project code uses TensorFlow/Keras APIs throughout the training and model definitions.

Results and documentation

• Results for FGSM experiments
• Results for random-noise experiments
• Additional training and analysis outputs
• Final presentation
• Final report
• Detailed expose

References

[1] Cisse, M., Bojanowski, P., Grave, E., Dauphin, Y., and Usunier, N. Parseval Networks: Improving Robustness to Adversarial Examples. 2017.

[2] Zagoruyko, S. and Komodakis, N. Wide Residual Networks. 2016.

Author

Sefika Efeoglu

Research Project, Data Science MSc, University of Potsdam