Provable and Measurable Machine Unlearning in Modern Learning Systems

Read full text Summarise with VIRAT
Authors: Chenglong Wang
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Machine learning systems often learn from large volumes of data, introducing privacy and security risks associated with persistent data influence. Machine unlearning removes the influence of specific training data from trained models, but retraining from scratch—the baseline for exact data removal—is prohibitively expensive for modern large-scale systems. Consequently, two key challenges arise: providing provable unlearning guarantees while preserving model utility and enabling efficient unlearning, and measuring unlearning completeness when such guarantees are unavailable. This thesis studies these challenges from a systems perspective, investigating how unlearning can be supported across three representative settings: structured learning systems, decentralized learning systems, and large-scale deep learning models. First, the thesis studies unlearning in structured systems through inductive graph learning. Training samples exhibit complex structural dependencies, and removing individual nodes may affect many parts of the learned representation. To address this challenge, the thesis proposes an efficient graph unlearning framework combining guided graph partitioning, subgraph repair, and similarity-based aggregation, enabling node-level deletion while preserving structural information and model utility. Second, the thesis studies provable unlearning under decentralized constraints through federated learning. It reduces how frequently clients and samples participate in model updates, lowering the effective retraining frequency when deletion requests occur. The thesis further introduces total variation stability to characterize how model updates change when a small number of clients or samples are removed. When sufficient stability holds, learning maintains reliable convergence even with fewer participants, enabling stable learning and efficient recovery of model utility after exact unlearning. Finally, recognizing that many practical methods are approximate and lack strict guarantees, the thesis introduces unlearning inference, which measures the residual influence of individual training samples in deployed models. It shows that existing membership inference–based evaluations are misaligned with this task and proposes a scalable framework for estimating sample-level unlearning completeness, revealing systematic risks of under-unlearning and over-unlearning. Together, these contributions connect system design, theoretical guarantees, and evaluation methods, advancing machine unlearning as a provable and measurable capability for controlling data influence in modern AI systems.

Fields of Study & Subjects

Privacy-Preserving Technologies in Data Adversarial Robustness in Machine Learning Cloud Data Security Solutions

Publication details

Publication date
1 January 2026
Language
EN
Document type
Thesis