README.md

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: Forget class accuracy(loss), Retain class accuracy(loss)

---

Training Details

Training Procedure

• Base Model: ResNet18
• Dataset: CIFAR-10
• Excluded Class: Varies by model
• Loss Function: Negative Log-Likelihood Loss
• Forget loss coefficient (alpha): 0.15
• Gradient normalization clip: 0.5
• Optimizer: SGD with:
  • Learning rate: 0.1
  • Momentum: 0.9
  • Weight decay: 5e-4
  • Nesterov: True
• Training Epochs: 1
• Batch Size: 2500
• Hardware: Single GPU (NVIDIA GeForce RTX 3090)

Algorithm

Loss Function for Unlearning

The overall loss function is defined as:

$$\mathcal{L}=\alpha \cdot {\mathcal{L}}_f+(1-\alpha )\cdot {\mathcal{L}}_r\backslash mathcal\{L\}\; =\; \backslash alpha\; \backslash cdot\; \backslash mathcal\{L\}\_f\; +\; (1\; -\; \backslash alpha)\; \backslash cdot\; \backslash mathcal\{L\}\_r$$

Gradient Update:

• Forget loss gradient ascent (negating gradients):

$$\theta \leftarrow \theta -\eta {\mathrm{\nabla }}_{\theta }{\mathcal{L}}_r+\eta \alpha {\mathrm{\nabla }}_{\theta }{\mathcal{L}}_f\backslash theta\; \backslash leftarrow\; \backslash theta\; -\; \backslash eta\; \backslash nabla\_\{\backslash theta\}\; \backslash mathcal\{L\}\_r\; +\; \backslash eta\; \backslash alpha\; \backslash nabla\_\{\backslash theta\}\; \backslash mathcal\{L\}\_f$$

• Gradient clipping:

$${\mathrm{\nabla }}_{\theta }\mathcal{L}\leftarrow \frac{{\mathrm{\nabla }}_{\theta }\mathcal{L}}{\max (1,\frac{\mathrm{\parallel }{\mathrm{\nabla }}_{\theta }\mathcal{L}\mathrm{\parallel }}{C})}\backslash nabla\_\{\backslash theta\}\; \backslash mathcal\{L\}\; \backslash leftarrow\; \backslash frac\{\backslash nabla\_\{\backslash theta\}\; \backslash mathcal\{L\}\}\{\backslash max(1,\; \backslash frac\{\backslash |\backslash nabla\_\{\backslash theta\}\; \backslash mathcal\{L\}\backslash |\}\{C\})\}$$

where ( C ) is the clipping threshold.

---

Model	Forget Class	Forget class acc(loss)	Retain class acc(loss)
cifar10_resnet18_AdvNegGrad_0.pth	Airplane	0.0 (28.448)	90.52 (0.631)
cifar10_resnet18_AdvNegGrad_1.pth	Automobile	0.0 (31.394)	91.27 (0.516)
cifar10_resnet18_AdvNegGrad_2.pth	Bird	0.0 (30.110)	92.72 (0.475)
cifar10_resnet18_AdvNegGrad_3.pth	Cat	0.0 (26.171)	92.44 (0.512)
cifar10_resnet18_AdvNegGrad_4.pth	Deer	0.0 (27.805)	91.19 (0.561)
cifar10_resnet18_AdvNegGrad_5.pth	Dog	0.0 (28.574)	92.81 (0.456)
cifar10_resnet18_AdvNegGrad_6.pth	Frog	0.0 (28.360)	92.18 (0.486)
cifar10_resnet18_AdvNegGrad_7.pth	Horse	0.0 (32.505)	92.89 (0.401)
cifar10_resnet18_AdvNegGrad_8.pth	Ship	0.0 (29.307)	91.34 (0.543)
cifar10_resnet18_AdvNegGrad_9.pth	Truck	0.0 (28.959)	92.47 (0.474)

---