How sisl-uri/deepcurer works
DeepCurer is a research workflow for creating backdoored classifiers and mitigating them with adversarial proxies and progressive neuron pruning. The principal path is: prepare data, poison and train a model, detect the target label and generate proxy examples, rank batch-normalization neurons, then prune while tracking clean accuracy and attack success. The README documents the workflow and supported attacks/datasets; source excerpts show key calls and implementations. Some orchestration is unsampled, so links supported only by the source index are identified as such through their evidence path.
Main entry points: deepcurer.py DeepCurer orchestration · backdoor_training.py Backdoor training · pruning.py Progressive pruning · train_test.py Model evaluation
Components 17
Data preparation
- CIFAR preparation
cifar10.py - Image datasets
Attack training
- Attack triggers
backdoor_triggers.py - Poisoned datasets
poison_dataset_generator.py - Backdoor training
backdoor_training.py
DeepCurer defense
- DeepCurer orchestration
deepcurer.py - Adversarial proxies
generating_adverserial_examples.py - Proxy dataset
- Neuron ranking
pruning_only.py - Progressive pruning
pruning.py - Pruning routines
pruning_only.py - Sanitized model
Model and evaluation
- ResNet models
resnet.py - VGG model
vgg.py - Model checkpoint
- Model evaluation
train_test.py
Other components
- Research user
Connections 21
- Research user → CIFAR preparation: prepares data
README.md - CIFAR preparation → Image datasets: organizes CIFAR-10
cifar10.py - Research user → Backdoor training: trains attack
README.md - Backdoor training → Attack triggers: selects trigger
backdoor_training.py - Backdoor training → Poisoned datasets: builds poisoned data
backdoor_training.py - Backdoor training → Image datasets: loads dataset
- Backdoor training → ResNet models: constructs classifier
backdoor_training.py - Backdoor training → VGG model: supports architecture
backdoor_training.py - Backdoor training → Model checkpoint: saves checkpoint
README.md - Research user → DeepCurer orchestration: runs defense
README.md - DeepCurer orchestration → Model checkpoint: loads checkpoint
deepcurer.py - DeepCurer orchestration → Adversarial proxies: generates proxies
deepcurer.py - DeepCurer orchestration → Attack triggers: uses attack trigger
deepcurer.py - DeepCurer orchestration → Poisoned datasets: builds test data
deepcurer.py - DeepCurer orchestration → Proxy dataset: creates proxy set
README.md - DeepCurer orchestration → Neuron ranking: requests ranking
deepcurer.py - Neuron ranking → Model evaluation: measures effects
pruning_only.py - DeepCurer orchestration → Progressive pruning: runs pruning
deepcurer.py - Progressive pruning → Model evaluation: evaluates candidates
pruning.py - Progressive pruning → Sanitized model: produces model
- DeepCurer orchestration → Model evaluation: tests classifier
deepcurer.py