federated unlearning

[mo-shahab]

In the provided script, the steps are outlined for federated unlearning, a process where a global model is trained collaboratively across multiple clients, and then specific client data is "forgotten" from the model to enhance privacy or security. Let's break down what each step does:

1. Federated Learning Settings:

   • This step sets the parameters for federated learning, such as the dataset to be used.

2. Client Data Loading and Model Initialization:

   • Data loaders for client training data and a common test set are loaded.
   • An initial global model is initialized.

3. Federated Learning and Unlearning Training:

   • The federated learning process begins. The global model is trained using data from selected clients. This involves multiple epochs of training.
   • After federated learning, unlearning processes are applied. This involves forgetting specific client data from the trained model. Two methods are mentioned:
     • Unlearning (FedEraser): Likely a specific unlearning method tailored for federated learning settings.
     • Unlearning without Calibration: Another method of unlearning, possibly a simpler variant.

4. Membership Inference Attack against Global Model (GM):

   • After training and unlearning, a membership inference attack (MIA) is conducted to assess the privacy of the trained models.
   • The attack model is trained based on the output of the target global model on client data loaders and test loaders.
   • The attack model's precision and recall are evaluated against different versions of the trained global model:
     • FL Standard: The original global model trained through federated learning.
     • FL Unlearn: The global model after unlearning specific client data.

The output includes precision and recall scores for the MIA against each version of the global model. Precision measures the proportion of correctly identified positives among all samples classified as positive by the attacker. Recall measures the proportion of correctly identified positives among all actual positives.

From the provided output:

• MIA Attacker Precision: The precision of the membership inference attack, indicating how accurately the attacker identifies membership in the dataset.
• MIA Attacker Recall: The recall of the membership inference attack, indicating how many actual members the attacker successfully identifies.

Comparing precision and recall between different versions of the global model (FL Standard and FL Unlearn) helps assess the effectiveness of the unlearning process in reducing the model's vulnerability to membership inference attacks and preserving privacy.

[mo-shahab]

woah, chill down g, too much effort, literally nobody read it