Detecting Usage Protocol Violations in Deep Learning Applications

This repo contains a replication package of the "Detecting Usage Protocol Violations in Deep Learning Applications" paper.

In the paper, we use two benchmarks:  

• The first benchmark we used from NeuraLint can be found under the director name Benchmark1/SOSamples
• The second benchmark we used from Humbatova et al. can be found under the director name Benchmark2/SOSamples

To reproduce the results in the paper:  

• Download the NeuralStateAnalysis Zip file.
• Extract the file and go to the NeuralStateAnlaysis directory.
• ( Optional ) Install the requirements by running 'Pip install requirements.txt.' N.B: The requirements.txt file is already in this repo.
  • To run NeuralState on Benchmark1:
    • Go to Benchmark1/SOSamples directory,
    • Open any of the programs you want to run,
    • Set the path: Path-to-folder/NeuralStateAnalysis/
    • Then, do 'python program_id.' Since the 'NeuralStateAnalysis(model).debug()' call is already present in all programs, you'll be able to reproduce results.
  • To run NeuralState on Benchmark2:
    • Go to Benchmark2/SOSamples directory,
    • Open any of the programs you want to run,
    • Set the path: Path-to-folder/NeuralStateAnalysis/
    • Then, do 'python program_id.' Since the 'NeuralStateAnalysis(model).debug()' call is already present in all programs, you'll be able to reproduce results.