CS-MIA: Membership inference attack based on prediction confidence series in federated learning.