Quantized models introduce discrete rounding operations into the computational graph, which may produce either genuine robustness against inference-time evasion attacks (coarser weight representation changing the decision boundary geometry) or gradient masking (rounding causing zero gradients that blind attacks). The current configuration evaluates pretrained TorchCV ResNet56, WRN-28-10, and DenseNet-100 models on CIFAR-100 across FP32, PTQ, and QAT variants using a layered attack suite. Dataset construction, preprocessing, class count, and TorchCV model identifiers are selected centrally in src/config.py.
To install dependencies
pip install -r requirements.txt
To download datasets (should be placed at root)
CIFAR-10
wget https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
curl -O https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
CIFAR-100
wget https://www.cs.toronto.edu/~kriz/cifar-100-python.tar.gz
curl -O https://www.cs.toronto.edu/~kriz/cifar-100-python.tar.gz
or set DATASET_DOWNLOAD=true in ./src/config.py
To run (and automatically generate visualizations)
python src/QuantAdv.py
To graph from incomplete results
python -m src.graphs.combine
To generate condensed figures for publication
python -m src.graphs.formaldata
Results are in ./data
