https://github.com/rkhosrowshahi/mormu
A novel multi-objective responsible machine unlearning using NSGA-II
https://github.com/rkhosrowshahi/mormu
evolutionary-algorithms machine-unlearning responsible-ai
Last synced: 4 months ago
JSON representation
A novel multi-objective responsible machine unlearning using NSGA-II
- Host: GitHub
- URL: https://github.com/rkhosrowshahi/mormu
- Owner: rkhosrowshahi
- Created: 2024-11-30T21:21:26.000Z (over 1 year ago)
- Default Branch: main
- Last Pushed: 2024-12-02T00:34:12.000Z (over 1 year ago)
- Last Synced: 2025-03-31T14:32:40.152Z (about 1 year ago)
- Topics: evolutionary-algorithms, machine-unlearning, responsible-ai
- Language: Python
- Homepage:
- Size: 45.9 MB
- Stars: 0
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Multi-Objective Responsible Machine Unlearning
In this project, we aim to propose a novel machine unlearning method for black-box neural networks such as Multi-layer Perceptron (MLP) and ConvNets (LeNet-5) using evolutionary multi-objective algorithm. We used non-dominated sorting genetic algorithm II (NSGA-II) []which is a meta-heuristic gradient-free algorithm to effectively optimize $`M \ge 2`$ objectives with respect to parameters such as weights and biases in neural networks.
The datasets used:
* Fashion-MNIST
There are two dataset splits in Fashion known as:
1. Training data $`D_{train}`$
2. Test data $`D_{test}`$
For the sake of unlearning, the training data is divided into two subsets as follows:
1. Forget data $`D_{u}`$
2. Retaining data $`D_{r}`$
The forgetting data is randomly selected from training data and the size is 1,000.
There are two networks to use and suggested commands are as follows:
1. MLP (64 hidden units), 50K params
```
python3 main.py --net mlp --dataset fashion --obj entropy_f1score --output_dir ./out/mlp-64 --steps 1000
```
2. LeNet-5, 60K params
```
python3 main.py --net lenet --dataset fashion --obj entropy_f1score --output_dir ./out/lenet --steps 1000
```