Read Me

摘要

Federated learning is an approach that ensures privacy in machine learning, but it has its limitations when it comes to preserving the right to be forgotten. To address this challenge, we propose a new method called Unlearning Key Revocation List (UKRL) for implementing federated unlearning. Our approach does not require clients' data or models to be unlearned; instead, we use revocation keys to remove clients from the model. We pre-trained the model to recognize these keys, so the model will forget the revoked clients when their revocation keys are applied. We conducted four experiments using MNIST datasets to verify the effectiveness of our approach, and the results showed that our work is not only effective but also time-saving since the unlearning time is 0. In conclusion, we provide a new perspective on achieving federated unlearning.

資料集：

訓練資料集

類別	資料集大小	data of revocation key	data	label
完整資料	120000筆	x_key_dtmp[i]	x_train_dtmp	y_train_dtmp
不包含類別9的資料	108102筆	x_key_dtmp[i]_without9	x_train_without9	y_train_without9
包含1000筆類別9的資料	109102筆	x_key_dtmp[i]_afew9	x_train_with_afew9	y_train_with_afew9

測試資料集

類別	資料集大小	data of revocation key	data	label
完整資料(不含金鑰)	10000筆	x_nokey_test	x_test	y_test
完整資料(含金鑰)	10000筆	x_key_test[i]	x_test	y_test
只包含類別9的資料(不含金鑰)	1009筆	x_nokey_test_9	x_test_9	y_test_9

• x_key1_test9: 只有用戶1的金鑰，測試只有包含類別9的資料，1009筆

• 1 <= i <= 用戶人數

下載連結：https://drive.google.com/file/d/1ky6WLjvpOERq3aWlEfIFGU5Mhg8PKof0/view?usp=sharing

程式碼

clients

檔名：doubleModel[i].py, 1<=i<=5

若有weight.txt檔，會將模型參數設成weight.txt的數字，再進行訓練。跑完以後，會把訓連完的參數存到 weight[i]_DbMdl.txt。

server

檔名：server_doubelModel.py

跑的時候，會讀取由clients的程式碼存下來的模型參數檔案weight[i]_DbMdl.txt，並把合併後的參數存到weight.txt。

執行

每個client分別執行後，執行server，得到global parameters。若global parameters測試結果未達訓練結束之標準，重複前述步驟。

需下載之套件：numpy, tensorflow, keras