User-Level Membership Inference for Federated Learning in Wireless Network Environment

Algorithm 2

Server’s attack procedure.

Input: The iteration round of GANs deployed on affiliated server , the federated learning model (), The iteration round of normal training , generator , discriminator , isolated model and victim’s updated parameters .

Output: The generated dataset :(x,y) and the inference result ‘IN’ or ‘OUT’.

Procedure Adversary Execution.

2: Initialize

Deceptively connect the affiliated server with to the victim

4: Set

for (;;) do

6: Train global model () by collecting all parameters from all participants

 Synchronize the following ‘for’ loop

8: Update () based on Eq (1)

end for

10: for () do

 Copy to affiliated server

12: Run to generate sample in a targeted manner

 Update based on Eq (8)

14: end for

16:

Output:

18: Output:

20: Attack Phase:

Train CNN model using dataset.

22: Perform membership inference attack against dataset.

Compare the inference results with the claimed information.

24: Output: Mark every record as ‘IN’ or ‘OUT’, where ‘IN’represents the Victim’s training sample.