Certified Machine Unlearning via Noisy Stochastic Gradient Descent

Authors: Eli Chien, Haoyu Wang, Ziang Chen, Pan Li

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that our approach achieves a similar utility under the same privacy constraint while using 2% and 10% of the gradient computations compared with the state-of-the-art gradient-based approximate unlearning methods for mini-batch and full-batch settings, respectively.
Researcher Affiliation Academia Eli Chien Department of Electrical and Computer Engineering Georgia Institute of Technology Georgia, U.S.A. ichien6@gatech.edu Haoyu Wang Department of Electrical and Computer Engineering Georgia Institute of Technology Georgia, U.S.A. haoyu.wang@gatech.edu Ziang Chen Department of Mathematics Massachusetts Institute of Technology Massachusetts, U.S.A. ziang@mit.edu Pan Li Department of Electrical and Computer Engineering Georgia Institute of Technology Georgia, U.S.A. panli@gatech.edu
Pseudocode Yes Algorithm 1 (Un)learning with PNSGD
Open Source Code Yes Our code is publicly available3. https://github.com/Graph-COM/SGD_unlearning
Open Datasets Yes We conduct experiments on MNIST [25] and CIFAR10 [26], which contain 11,982 and 10,000 training instances respectively.
Dataset Splits No The paper does not explicitly specify a separate validation dataset split or how it was used to tune hyperparameters or monitor training progress.
Hardware Specification Yes The codes run on a server with a single NVIDIA RTX 6000 GPU with AMD EPYC 7763 64-Core Processor.
Software Dependencies Yes All the experiments run with Py Torch=2.1.2 [36] and numpy=1.24.3 [37].
Experiment Setup Yes We set the learning iteration T = 10, 20, 50, 1000 to ensure PNSGD converges for mini-batch size b = 32, 128, 512, n respectively. All results are averaged over 100 independent trials with standard deviation reported as shades in figures. We set the step size η for the PNSGD unlearning framework across all the experiments as 1/L.