Federated Unlearning

This topic sits at the intersection of two concepts you already know: Machine Unlearning (removing the effect of a specific record from a trained model) and Federated Learning (training a model across many devices without collecting their raw data on a central server). Understanding federated unlearning is crucial for building real-world products where users demand privacy. When a user asks to be deleted, your system must actually remove their influence from the model, even though you never stored their personal data in one place. This is a core requirement for Data Privacy Compliance and AI Governance. Regulators are increasingly asking: 'If a citizen demands erasure, can your technology truly comply?' Without federated unlearning, the answer is no, leaving both legal and technical failures.

Imagine a classroom where every student keeps their own diary. The teacher never sees the diaries. Instead, at the end of each week, every student writes a short 'summary note'—just a few lines capturing what they learned—and hands those notes to the teacher. The teacher blends all these notes into a single 'class summary,' which becomes the starting point for the next week. The teacher only ever sees the summaries, never the original diaries. This is federated learning: each device trains on its own data and sends only a small update (the summary) to the central server, which averages them to improve the global model.

Now, suppose one student leaves and says, 'Please remove everything I contributed from the class summary.' The teacher faces a problem: they never had access to that student’s diary, so they cannot simply delete the original entry. They only have the summary notes they collected. Their influence has been blended into the class summary over time, much like a drop of ink in a glass of water. You cannot pick the ink out; it has already soaked through the water.

The central difficulty is that each week’s summary was built on top of previous summaries. The leaving student’s influence is embedded in the entire model structure. To be absolutely certain their influence is gone, you must throw away the entire class summary and start over from zero, asking every remaining student to write new summary notes as if nothing had happened before. This 'full retraining' is correct but extremely expensive and slow because it requires contacting every device again.

To make this efficient, we change how the classroom operates. Imagine the class is split into two separate groups (let us call them Group A and Group B). The teacher maintains two separate class summaries: one for Group A and one for Group B. When a student joins, they are assigned to exactly one group and only contributes to that specific summary. If a student from Group A leaves, the teacher does not need to touch the Class Summary for Group B. They simply destroy the Class Summary for Group A and retrain it using only the remaining students in Group A. This method of dividing data into non-overlapping groups so that the removal of one member only requires retraining a small fraction of the total model is called SISA-style sharded training.

Think of a restaurant kitchen that prepares soups for large banquets. In a standard setup, all chefs add their spice blends to one giant pot. If one chef leaves and demands their recipe be forgotten, the head chef must dump the entire giant pot and start over with all the remaining chefs—a massive waste of time and ingredients.

To solve this, imagine the kitchen now has several separate, identical small pots. Each chef is assigned to stir only one specific pot for the entire duration. The soups in these pots are kept completely distinct; no spice from Pot 1 ever touches Pot 2. If Chef Smith, who was stirring Pot 1, leaves and asks for their influence to be removed, the head chef simply dumps out Pot 1. Pot 2 remains perfectly untouched and ready to serve. The head chef then re-cooks only Pot 1 with the remaining chefs assigned to it.

This separation ensures that removing one participant’s influence requires restarting only a tiny fraction of the total work. In federated unlearning, this division of labor allows systems to comply with deletion requests almost instantly for the affected part, while leaving the rest of the system running smoothly.

Federated unlearning is the process of removing a specific participant's data influence from a machine learning model trained across distributed devices without central raw data access, typically achieved by using independent subsets of data so that deleting one member only requires retraining a small subset of the model rather than the entire system.

SISA-style sharded training is a method where the training dataset is divided into non-overlapping groups (shards), ensuring that no single piece of data belongs to more than one group, allowing independent updates and deletions within each shard.

This concept relies on your notes about Machine Unlearning and Federated Learning. Machine Unlearning provides the goal: erasing specific data's effect. Federated Learning provides the constraint: no central raw data. This topic bridges them by introducing a structural solution (SISA-style sharded training) that makes unlearning feasible within the federated constraint. It also connects directly to Deletion Requests and the Right to Be Forgotten, as these legal frameworks mandate that models must be able to 'forget' individuals, which federated unlearning technically enables.