When to Forget? Complexity Trade-offs in Machine Unlearning

Authors: Martin Van Waerebeke, Marco Lorenzi, Giovanni Neglia, Kevin Scaman

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Experiments: We investigate the global landscape of the unlearning complexity ratio as a function of key factors, including the accepted excess risk threshold e and the unlearning budget (ϵ, δ), which are jointly quantified by the constant κϵ,δ. 5.1. Experimental Setting: The goal of the experiment section is to validate the theoretical analysis presented in Section 4 by comparing the performance of unlearning and retraining on both real and synthetic functions and datasets. 5.2. Experiments on Synthetic Data. 5.3. Experiments on Real Data. 5.4. Experimental Results: Figure 2 illustrates the empirical unlearning complexity ratio.
Researcher Affiliation Academia 1INRIA Paris 2INRIA Sophia Antipolis. Correspondence to: Martin Van Waerebeke <EMAIL>.
Pseudocode Yes Algorithm 1 Iterative (Un)Learning Algorithm Algorithm 2 Noise and Fine-Tune Unlearning Algorithm
Open Source Code No The paper does not contain any explicit statements about releasing code or links to source code repositories for the methodology described.
Open Datasets Yes 5.1. Experimental Setting: We aim to learn linear regression models in Rd (with even d). We perform experiments both on synthetic worst-case functions, as analysed in our theory, and on the Digit dataset of handwritten digits, which is a subset of the larger dataset proposed in Alpaydin and Alimoglu (1996). Alpaydin, E. and Alimoglu, F. (1996). Pen-Based Recognition of Handwritten Digits. UCI Machine Learning Repository. DOI: https://doi.org/10.24432/C5MG6K.
Dataset Splits Yes In every experiment, the retain and forget are obtained through the random splitting of the dataset into two parts of respective sizes n rf n and rf n . We consider rf = 10 2.
Hardware Specification No The paper mentions experiments being performed but does not provide specific details about the hardware used (e.g., GPU models, CPU models, memory specifications).
Software Dependencies No The paper mentions using 'stochastic gradient descent' and 'the standard SGD optimizer' but does not provide specific version numbers for any software libraries, frameworks, or programming languages.
Experiment Setup Yes 5.3. Experiments on Real Data: For the real data, the experimental process is simpler as we optimize a standard cross-entropy loss with L2 regularization. For various values of κϵ,δ and e, we measure T U e and T S e in a more realistic machine learning setting, with decaying learning rate, batch size of 64, and averaging the results over 50 runs. The learning rate is initialised at 10 2 and multiplied by 0.6 every 1000 epoch. Each experiment is repeated 50 times and the results are then averaged. We used a batch-size of 64 and trained until the threshold e was reached, for every chosen value of κϵ,δ. The optimized used is the standard SGD optimizer without acceleration.