Quantum Machine Unlearning Research Code

This repository contains code for machine unlearning experiments, including:

1. Data poisoning and visualization
2. Hessian-based analysis of model landscapes
3. Various unlearning methods comparison (Retrain, Finetune, Scrub, Grad-Asc)
4. Evaluation of unlearning effectiveness on both classical MLP and quantum neural networks

The paper reference is https://arxiv.org/abs/2508.02422 where we for the first time proposed the concept of Quantum Machine Unlearning.

File Structure

• datar.py: Data preprocessing and loading functions
• models.py: Model for traning and unlearning definitions for both MLP and QNN
• poison_unlearn.py: Implementation of various unlearning method wrappers
• run_unlearn_mnist.py: Main script for MNIST unlearning experiments
• run_unlearn_xxz.py: Main script for XXZ model unlearning experiments
• run_example.py: Run the code with only data poisoning part
• run_test.py: A simplified test script for quick verification on functionality parallel run_unlearn_mnist.py
• plot_*.ipynb: Visualization related notebooks

Requirements

Install all requirements with (python 3.12 suggested):

pip install -r requirements.txt

This work is enabled by the high performance quantum-classical hybrid software infrastructure of TensorCircuit-NG.

Usage

Quick Start

Run the example script to test the code:

python run_example.py

Full Experiments

To run the MNIST unlearning experiments:

python run_unlearn_mnist.py

To run the XXZ model unlearning experiments:

python run_unlearn_xxz.py

Notes

• No specified hardware is required.

• The typical environment setup and software installation time is around 10 minutes.

• Part of the expected output from the demo run_example.py (expected runtime on a desktop: ~5min):

Machine Unlearning Examples
========================================

1. MNIST Example:
Loading MNIST data...
Training: (500, 784)
250 250
Validation: (1000, 784)
Running label flipping experiment with QNN...
F=5.71 C=6.53 S=12.00 P=13.19: 100%|████████████| 64/64 [00:01<00:00, 43.69it/s]

Results: [{'depth': 4, 'epochs': 10, 'learning_rate': 0.005, 'batch_size': 256, 'acc': Array(0.624, dtype=float32), 'val_acc': Array(0.591, dtype=float32)}, {'depth': 4, 'epochs': 10, 'learning_rate': 0.005, 'batch_size': 256, 'acc': Array(0.67, dtype=float32), 'val_acc': Array(0.69100004, dtype=float32)}, {'depth': 4, 'epochs': 10, 'learning_rate': 0.005, 'batch_size': 256, 'acc': Array(0.70400006, dtype=float32), 'val_acc': Array(0.744, dtype=float32)}, {'depth': 4, 'epochs': 10, 'learning_rate': 0.005, 'batch_size': 256, 'acc': Array(0.57000005, dtype=float32), 'val_acc': Array(0.67200005, dtype=float32)}, {'depth': 4, 'epochs': 10, 'learning_rate': 0.005, 'batch_size': 256, 'acc': Array(0.476, dtype=float32), 'val_acc': Array(0.47200003, dtype=float32)}, {'depth': 4, 'epochs': 10, 'learning_rate': 0.005, 'batch_size': 256, 'acc': Array(0.54200006, dtype=float32), 'val_acc': Array(0.513, dtype=float32)}, {'depth': 4, 'epochs': 10, 'learning_rate': 0.005, 'batch_size': 256, 'acc': Array(0.522, dtype=float32), 'val_acc': Array(0.48700002, dtype=float32)}, {'depth': 4, 'epochs': 10, 'learning_rate': 0.005, 'batch_size': 256, 'acc': Array(0.59400004, dtype=float32), 'val_acc': Array(0.34, dtype=float32)}, {'depth': 4, 'epochs': 10, 'learning_rate': 0.005, 'batch_size': 256, 'acc': Array(0.614, dtype=float32), 'val_acc': Array(0.27100003, dtype=float32)}, {'depth': 4, 'epochs': 10, 'learning_rate': 0.005, 'batch_size': 256, 'acc': Array(0.67800003, dtype=float32), 'val_acc': Array(0.261, dtype=float32)}]

2. XXZ Model Example:
Loading XXZ data...

Methods

The repository implements several machine unlearning methods:

1. Retrain: Retrain the model from scratch on the retained data
2. Fine-tune (cf): Continue training on the retained data
3. Scrub: Use KL divergence regularization to "scrub" the influence of forgotten data
4. Gradient Ascent (ga): Use gradient ascent to actively unlearn the forgotten data

Models

Two types of models are implemented:

1. Classical MLP: Multi-layer perceptron with configurable hidden layers
2. Quantum Neural Network (QNN): Quantum circuit-based model using TensorCircuit

Data

The code supports experiments on:

1. MNIST: Binary classification of digits 1 and 9
2. XXZ Model: Quantum many-body system data (requires external dataset files)

Results

The experiments evaluate:

1. Validation Accuracy: Performance on clean validation data
2. Forgetting Accuracy: Ability to forget the "unlearned" data
3. Model Stability: Robustness of the unlearning process

License

This project is licensed under the MIT License - see the LICENSE file for details.