README.md

All the files in this folder need to be inside Marabou's root directory to work. Instructions for building marabou are given here. Each of the experiments in the report is described below:

1. Testing robustness on regions using explicit constraints:

    python testmarabouwithl1constraints.py
   

   To test it with your own defined regions, you can edit the main method, or import it and call check(<path to your image>, <delta>, <l1norm>, <region>). Note that going above 3x4 pixel regions might give memory issues.

2. Testing robustness on regions using auxilliary networks:

    python testmarabouregions_v2.py
   

   Or import the file and call check(<path to your image>, <delta>, <l1norm>, <region>)

3. Robustness on entire image (L-1/clipped norms):

    python testmarabouonl1_v2.py --delta <delta> --l1norm <l1norm>
   

4. Vanilla L-infinity robustness:

    python testmarabou.py --delta <delta>
   

The image recored on SAT is stored in Marabou/recovery/ do NOT delete the folder. The program only takes onnx networks as inputs. We have provided two example networks with our code. For generating new networks, please execute l1net_generator.ipynb. We also provide an inference tester for onnx networks.

Note: we provide example images, but you will have to edit test_utils.py for new images.