How ongwinkent/maverick works
Maverick is a medical-imaging federated-learning and federated-unlearning research framework. Its training workflow loads MedMNIST data, partitions retain and unlearn examples across clients, performs local updates with federated aggregation, evaluates the global model, and can save it. A separate unlearning entry point loads a trained model, applies a selected method—including local Maverick sensitivity optimization or baseline methods—and evaluates retain accuracy, unlearn accuracy, and membership-inference risk. The graph follows shown calls and README capabilities; unsampled implementation details are not expanded.
Main entry points: training.py Federated trainer · unlearn_func.py Unlearning workflow · fl_main.py FL entry point · unlearn_main.py Unlearning entry point
Components 15
Entry and configuration
- FL entry point
fl_main.py - Training options
configs/train_config.py - Unlearning entry point
unlearn_main.py - Unlearning options
configs/unlearn_config.py
Federated training
- Federated trainer
fl_strategies/training.py - Client updates and aggregation
fl_strategies/utils.py
Unlearning
- Unlearning workflow
unlearn_strategies/unlearn_func.py - Unlearning methods
unlearn_strategies/strategies.py
Data, models and evaluation
- MedMNIST and partitions
src/dataset.py - Model constructors
model/models.py - ResNet architectures
model/resnet.py - Evaluation and privacy metrics
src/metrics.py - Runtime utilities
src/utils.py
Other components
- Researcher
- MedMNIST
Connections 20
- Researcher → FL entry point: starts training
fl_main.py - FL entry point → Training options: reads options
fl_main.py - FL entry point → Runtime utilities: sets seed
fl_main.py - FL entry point → Federated trainer: starts trainer
fl_main.py - Federated trainer → MedMNIST and partitions: loads and partitions
fl_strategies/training.py - MedMNIST and partitions → MedMNIST: loads datasets
src/dataset.py - Federated trainer → Model constructors: constructs model
fl_strategies/training.py - Model constructors → ResNet architectures: builds ResNet
model/models.py - Federated trainer → Client updates and aggregation: selects and trains clients
fl_strategies/training.py - Federated trainer → Evaluation and privacy metrics: evaluates global model
fl_strategies/training.py - Federated trainer → Runtime utilities: configures device
fl_strategies/training.py - Researcher → Unlearning entry point: starts unlearning
unlearn_main.py - Unlearning entry point → Unlearning options: reads options
unlearn_main.py - Unlearning entry point → Unlearning workflow: runs workflow
- Unlearning workflow → MedMNIST and partitions: loads and partitions
unlearn_strategies/unlearn_func.py - Unlearning workflow → Model constructors: constructs model
unlearn_strategies/unlearn_func.py - Unlearning workflow → Unlearning methods: dispatches method
unlearn_strategies/unlearn_func.py - Unlearning methods → Client updates and aggregation: uses optimization helpers
- Unlearning workflow → Evaluation and privacy metrics: evaluates unlearned model
unlearn_strategies/unlearn_func.py - Unlearning workflow → Runtime utilities: configures device
unlearn_strategies/unlearn_func.py