What is Machine Unlearning?

Machine Unlearning for Dummies

Say machine learning is a machine learning from data, then machine unlearning is a machine forgetting the knowledge that it has learnt from data.

To be more exact, we want to erase a machine’s knowledge that it has learnt from certain data, making it look like that the machine has never seen that data.

In a way, you can say we are rewriting history for the machine, or the machine has got amnesia that has caused permanent memory loss of certain data.

If you like the analogy using the human brain, machine unlearning is like memory erasure, which is very difficult.

MIB

IRL, we don’t have MIB’s neuralyzer, so…

A Little Bit of History

There have been discussions of decremental learning early in the 00s, such as Cauwenberghs and Poggio in NeurIPS 2000.

In 2015, Cao and Yang (Talk) first put together the term “machine unlearning”. (Fun fact: HKUCS’s Dr. H. Cui got his PhD under Yang’s supervision at Columbia) The paper appeared in the 2015 IEEE Symposium on Security and Privacy (“Oakland”), a top conference in cybersecurity and privacy (CORE A* rank).

In 2019, Bourtoule et al. proposed SISA training, a framework that expedites the machine unlearning process. The paper later appeared in IEEE S&P 2021.

In 2022, Nguyen et al. comprehensively surveyed recent developments of the machine unlearning field.

What is Federated Learning?

Federated Learning in Google’s Own Words

It works like this: your device downloads the current model, improves it by learning from data on your phone, and then summarizes the changes as a small focused update. Only this update to the model is sent to the cloud, using encrypted communication, where it is immediately averaged with other user updates to improve the shared model. All the training data remains on your device, and no individual updates are stored in the cloud. – McMahan et al. (Paper)

Quantitative Performance and Security Evaluation of Federated Learning on open-sourced platforms by C.T. Fong – also supervised by Prof. S.M. Yiu, an industry project with ASTRI (I spent a summer there. ASTRI’s CTO Dr. Lucas C.K. Hui was with HKUCS).

About This Project

This project aims to investigate how machine unlearning can be applied in federated learning, to produce materials for non-specialists and experts in related areas for learning this topic, to explore possible improvements to the work done in the field and to deliver final year project results of outstanding quality and standard.

Related resources in security of ML and machine unlearning.