CRhapsody / nnrepair

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

NNRepair: Constraint based Repair of Neural Network Classifiers

Directories

Neural Network Models

Neural Network models used in paper are stored under NN-Code directory. This directory contains model for MNIST-LowQuality, MNIST-Poisoned, MNIST-Adversarial, Cifar10-Poisoned and Cifar10-Adversarial.

Code for combining experts

The code to combine experts is located inside CombinationCode directory.

Repair constraints generated using SPF

The constraints file are located in Constraints directory. This directory has 5 subdirectories i.e., MNIST-LowQuality, MNIST-Poisoned, MNIST-Adversarial, Cifar10-Poisoned and Cifar10-Adversarial. Within each models directory, we have grouped the constraint files by Intermediate-Layer Repair and Last-Layer Repair. The constraints file names follow the naming convention repairFor_label.txt (e.g., repairFor1 refers to constraint file for label 1). Please note that we had 4 different experiments for each repair (0, 10, 50 and 100 passing tests). These 4 scenarios are named as ExpA, ExpB, ExpC and ExpD respectively.

Z3 Solutions for Constraint Files

The constraints file are located in Z3Solutions directory. This directory has 5 subdirectories i.e., MNIST-LowQuality, MNIST-Poisoned, MNIST-Adversarial, Cifar10-Poisoned and Cifar10-Adversarial. Within each models directory, we have grouped the solution files by Intermediate-Layer Repair and Last-Layer Repair. The solution file names follow the naming convention solution_label.txt (e.g., solution1 refers to solution file for label 1). Please note that we had 4 different experiments for each repair (0, 10, 50 and 100 passing tests). These 4 scenarios are named as ExpA, ExpB, ExpC and ExpD respectively.

Final Results

The results file are located in Results directory. This directory has 5 subdirectories i.e., MNIST-LowQuality, MNIST-Poisoned, MNIST-Adversarial, Cifar10-Poisoned and Cifar10-Adversarial. Within each models directory, we have grouped the results files by Intermediate-Layer Repair and Last-Layer Repair. The results file names follow the naming convention model_repairtype_Exp#_dataset.csv (e.g., POISONED_CIFAR_LAST_LAYER_ExpA_POISONED_TEST.csv refers to results on Poisoned Test Dataset when last layer repair was performed on CIFAR POISONED MODEL). Please note that we had 4 different experiments for each repair (0, 10, 50 and 100 passing tests). These 4 scenarios are named as ExpA, ExpB, ExpC and ExpD respectively.

About


Languages

Language:Java 100.0%