metadata
license: mit
datasets:
- uoft-cs/cifar10
language:
- en
metrics:
- accuracy
- confusion_matrix
base_model:
- jaeunglee/resnet18-cifar10-unlearning
tags:
- machine_unlearning
- classification
Evaluation Report
Testing Data
Dataset: CIFAR-10 Test Set
Metrics: Top-1 Accuracy
Training Details
Training Procedure
- Base Model: ResNet18
- Dataset: CIFAR-10
- Excluded Class: Varies by model
- Loss Function: CrossEntropyLoss
- Optimizer: SGD with:
- Learning rate: 0.1
- Momentum: 0.9
- Weight decay: 5e-4
- Nesterov: True
- Scheduler: CosineAnnealingLR (T_max: 200)
- Training Epochs: 200
- Batch Size: 128
- Hardware: Single GPU (NVIDIA GeForce RTX 3090)
Algorithm
The AdvNegGrad algorithm was employed for targeted unlearning. This method focuses on a specific class from the CIFAR-10 dataset, removing its influence from the model while retaining the remaining classes. Each resulting model (cifar10_resnet18_AdvNegGrad_X.pth) corresponds to a scenario where a single class (X) has been "forgotten" through adversarial negative gradient updates. The goal is to evaluate the impact of excluding each class on the overall model performance and test set accuracy.
Results
| Model | Excluded Class | CIFAR-10 Accuracy (%) |
|---|---|---|
| cifar10_resnet18_AdvNegGrad_0.pth | Airplane | 33.97 |
| cifar10_resnet18_AdvNegGrad_1.pth | Automobile | 33.72 |
| cifar10_resnet18_AdvNegGrad_2.pth | Bird | 37.70 |
| cifar10_resnet18_AdvNegGrad_3.pth | Cat | 44.12 |
| cifar10_resnet18_AdvNegGrad_4.pth | Deer | 37.75 |
| cifar10_resnet18_AdvNegGrad_5.pth | Dog | 37.62 |
| cifar10_resnet18_AdvNegGrad_6.pth | Frog | 44.38 |
| cifar10_resnet18_AdvNegGrad_7.pth | Horse | 38.20 |
| cifar10_resnet18_AdvNegGrad_8.pth | Ship | 30.38 |
| cifar10_resnet18_AdvNegGrad_9.pth | Truck | 27.55 |
Notes
- The Top-1 Accuracy metric represents the percentage of correctly classified samples from the CIFAR-10 test set.
- The excluded class refers to the class omitted during model training to evaluate its effect on accuracy.
- Results for additional models are pending computation.
Conclusion
This report provides a structured comparison of CIFAR-10 accuracy across models with different excluded classes. Further analysis is required to determine the impact of each excluded class.