chengtang48 / adversarial_lower_bound Goto Github PK

Implementation and tests for original IBP methods (in matrix form), with support of general linear layers (simple linear, conv2d, mean/max pooling), and with support of batch processing (equivalent to IBP-Lin with null schedule vector; see Algo 1 in paper)

• fast_appx.py
• utils.py

Implementation and tests for IBP-Lin (see Algorithm 1 in paper for pseudo code)

• linear_appx.py

Architecture, training, and loading functions of neural nets used for experiments

• relu_networks.py (small ReLU nets)
• medium_size_networks.py (LeNet-5)
• pretrained_imagenets.py (VGG imagenets)

1. Tested IBP with small, medium, and pretrained imagenet models
2. Tested IBP-Lin with small model

1. Sanity check of theoretical findings on feed-forward ReLU networks

   • for random gaussian networks, understand how does the hidden dimension of the network affect the performance of k-layer linear approximation performance in IBP-Lin (Prop 3.2)

   • demonstrate how the quality of approximation can change sharply with magnitude of network weights (Prop. 3.4)

   • run IBP-Lin with increased k in k-layer approximation doesn't necessarily outperform the original IBP bounds (Prop. 3.5)

2. Additional implementation and tests

   • add support of pool layers for IBP-Lin (in linear_appx.py)
   • more tests for conv2d/pool layers for IBP-Lin