NeuPerm

Code for the paper: NeuPerm: Disrupting Malware Hidden in Neural Network Parameters by Leveraging Permutation Symmetry.

License

This work is based on a patent-pending request, and is licensed for non-commerical non-derivative use only.
See the License.

Setup

• Create a conda Python environment: conda create --channel-priority flexible -n neuperm -f environment.yaml.
• Set IMAGENET12_ROOT in neu_perm/config.py to a directory that contains the files of the ImageNet12 validation dataset.

Running Experiments

Experiment 1: Model Performance

• Configure run parameters in experiments/exp1.py:
  • model_names: List of model names; options are: ['densenet121','resnet50','resnet101','vgg11','llama-3.2-1b']
  • prune_amounts: List of floats between 0 and 1; what fraction of the parameters to prune.
  • epsilons: List of floats to use as the scale of the normal random noise.
  • n_repeats: Positive integer. number of times to repeat experiments.
  • device: Either 'cpu' or 'cuda'. What device to run models on.
  • batch_size: Batch size for inference.
• Run the script:
  • (Recommended) run detached from terminal: nohup python -u experiments/exp1.py > out_exp1.txt

Experiment 2: NeuPerm Disrupts MaleficNet

Since this is based on replicated MaleficNet models, I won't publish the code for this experiment as it's based on code that's not mine.