MASIMU: Multi-Agent Speedy and Interpretable Machine Unlearning
Keywords: multi-agent, unlearning, interpretable, faster, robust, MASIMU, LIME, reinforcement learning, explainable AI, XAI
TL;DR: Multi-Agent Speedy and Interpretable Machine Unlearning is a novel AI privacy framework for sensitive information, which is faster with multiple agents reducing the observation space per agent and interpretable with LIME XAI method.
Abstract: The regulatory landscape around the use of personal data to train AI/ML models is rapidly evolving to protect privacy of sensitive information like user locations or medical data and improve AI trustworthiness. Practitioners must now provide the capability to unlearn or forget data---the forget set---that was used to train an AI model, without triggering a full model re-train on the remaining data---the retain set to be computationally efficient. Existing unlearning approaches train via some combination of fine-tuning pre-trained AI models solely on the retain set, pruning model weights then unlearning, and model-sparsification-assisted unlearning. In our research paper, we use deep learning (DL), multi-agent reinforcement learning (MARL) and explainable AI (XAI) methods to formulate a faster, more robust and interpretable unlearning method than past works. Our method, multi-agent speedy and interpretable machine unlearning (MASIMU), fine-tunes a pre-trained model on the retain set, interpretably re-weighting the gradients of the fine-tuned loss function by computing the similarity influences of the forget set on the batched retain set based on weights generated by an XAI method. We add a MARL framework on top to address the challenge of high dimensional training spaces by having multiple agents learning to communicate positional beliefs and navigate in image environments. The per-agent observation spaces have lower dimensions, leading to the agents focusing on unlearning interpretable gradients of important superpixels that influence the target labels in the learning criteria. We provide extensive experiments on four datasets---CIFAR-10, MNIST, high resolution satellite images in RESISC-45, skin cancer images in HAM-10000 to unlearn for preserving medical privacy---computing robustness, interpretability, and speed relative to the dimensionality of the training features, and find that MASIMU outcompetes other unlearning methods.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13057
Loading