Thesis-MUL: Saliency-Based Machine Unlearning

This repository implements and evaluates saliency-based machine unlearning (SalUn) with a focus on two experiment families:

• Class-wise forgetting: remove specific semantic classes (cats, dogs, vehicles).
• Random-data forgetting: remove random subsets (10/20/30%) of Tiny ImageNet.

The codebase pairs quantitative evaluation with feature visualization (Lucent) to verify whether knowledge was actually erased inside the network.

Quick Start

conda env create -f environment.yml
conda activate salun

# Typical data layout
datasets/tiny-imagenet-200/

# Run a random-data forgetting notebook
jupyter notebook notebooks/random_data/20_random_data.ipynb

Repository Map (organized by forgetting mode)

experiments/
├── class_wise/
│   ├── masks/            # Saliency masks per class family
│   ├── models/           # Class-wise checkpoints
│   ├── indices/          # Forget-set indices & masks
│   ├── scripts/          # Training/unlearning entrypoints
│   └── notebooks/        # (reserved) class-wise notebooks
├── random_data/
│   ├── masks/            # Masks for random 10/20/30% forgetting
│   ├── results/          # Quantitative + good_results
│   ├── forgotten_images/ # Exported forgotten samples
│   ├── scripts/          # Pipelines for random forgetting
│   └── notebooks/        # (reserved) random-data notebooks
└── common/
   └── resnet50_pretrained.pth  # Base model

notebooks/
├── random_data/          # 10/20/30_random_data, comparison_experiments, config.py
├── class_wise/           # Resnet18/, resnet-50/, inceptionv3/, legacy/
└── common/               # diversity, feature_inversion, modelzoo, neuron_interaction, tutorial

src/                      # Training, data, utils
scripts/                  # Analysis helpers (forgotten data, visualization)
labels/, datasets/        # Tiny ImageNet labels/data
models/, results/, visuals, plots  # Outputs and figures

Key Entry Points

• Random-data forgetting notebooks: notebooks/random_data/10_random_data.ipynb, 20_random_data.ipynb, 30_random_data.ipynb, comparison_experiments.ipynb
• Class-wise notebooks: notebooks/class_wise/resnet-50/…, notebooks/class_wise/Resnet18/…, notebooks/class_wise/inceptionv3/… (legacy variants kept in notebooks/class_wise/legacy/)
• Scripts (class-wise): experiments/class_wise/scripts/resnet50_unlearn_*.py, generate_all_masks.py, generate_mask_class.py, main_random_class.py
• Scripts (random-data): see experiments/random_data/scripts/ (add new runs here if needed)
• Base model: experiments/common/resnet50_pretrained.pth

Data & Models

• Dataset: Tiny ImageNet under datasets/tiny-imagenet-200/
• Forget-set metadata: experiments/class_wise/indices/*.pt
• Masks: experiments/class_wise/masks/ and experiments/random_data/masks/
• Checkpoints: experiments/class_wise/models/ plus base model in experiments/common/
• Results: experiments/random_data/results/good_results/ contains RL and conservative variants

Workflow

1. Choose forgetting mode

• Class-wise: run scripts in experiments/class_wise/scripts/
• Random-data: run notebooks in notebooks/random_data/

2. Train/unlearn

• Use SalUn with random-label (RL) variants (see notebook configs)

3. Validate

• Quantitative: accuracy on retain vs forget splits (stored in results)
• Qualitative: Lucent activation grids and neuron visualizations (common notebooks)

Notes

• Large artifacts are ignored by .gitignore; keep big checkpoints under experiments/ or models/.
• CLEANUP_GUIDE.md and CLEANUP_SUMMARY.md describe how the repo was reorganized.
• Legacy class-wise notebooks are preserved under notebooks/class_wise/legacy/ for reference.

License

MIT License. See LICENSE.