Revnets

Reverse engineer internal parameters of black box neural networks

Usage

Run

revnets

The entry point of this command is the main function in revnets.main.main. This command will start an experiment to recover the weights of a trained neural networks. Each experiment consists of three configurable components:

1. pipelines: architecture + dataset that produces target networks that will be recovered
   • All possible networks are defined in revnets.pipelines
   • The networks used in the experiment are configured with the option network in the config file
2. reconstructions: the techniques used to recover neural network weights
   • All possible techniques are defined in revnets.reconstructions
   • The networks used in the experiment are configured in the init file
3. evaluations: the methods used to evaluate a weight reconstruction
   • All possible evaluations are defined in revnets.evaluations
   • The evaluations used in the experiment are configured in the init file

Hyperparameters and other options are specified in a config file located at config.yaml

An example config file is provided at config.yaml

For all possible options, see config.py

Installation

pip install git+https://github.com/quintenroets/revnets.git

Installation for development

Clone the project and run

pip install -e .