GBDTC Team Wins Distinguished Paper Award at ACM CCS

Redefining Machine Unlearning in Collaborative AI Systems

Split Unlearning, authored by Yanna Jiang, Saber Yu, Xu Wang, Baihe Ma, and Ren Ping Liu (GBDTC), in collaboration with Wei Ni and Qin Wang (Data61, CSIRO), introduces a novel unlearning framework tailored for Split Learning (SL)—a collaborative learning paradigm where clients and servers train different parts of a model without sharing raw data. Split Unlearning enables efficient and privacy-preserving removal of user data from SL models, achieving 0% accuracy on forgotten labels while improving retained performance by 8%.

Responding to Privacy Laws: GDPR, Privacy Act 1988

As AI becomes embedded in everyday decision-making, ensuring individuals can exercise their “right to be forgotten”—as defined in the EU’s General Data Protection Regulation (GDPR) and Australia’s Privacy Act 1988—is increasingly critical. However, most machine learning systems, especially collaborative frameworks like SL, lack efficient mechanisms for compliant unlearning.

Split Unlearning offers a practical and scalable response to this challenge. It enables models trained under Split Learning frameworks to forget specific user data—even when the users are no longer online or accessible—while preserving the integrity and performance of the remaining system. This represents a major step forward in making privacy-compliant AI both technically and operationally feasible.

Measurable Impact and Real-World Applications

Split Unlearning demonstrates strong performance across diverse data distributions and tasks, effectively removing targeted user information while preserving the overall utility of the model. Its low and consistent communication and computation overhead makes it especially suitable for real-world deployment.

Potential application scenarios include:

•             Healthcare AI: Removing patient data from diagnostic models upon consent withdrawal
•             Financial services: Ensuring client trace data is erasable from investment prediction models
•              Edge computing: Supporting fast local forgetting on resource-constrained devices
•             Cross-organisational collaboration: Empowering federated learning with trust and regulatory compliance