Differential Privacy with Varying Epsilon Values in parameter: No Impact on Output

[Hafizamariaiqbal]

🐛 Bug

Using Differential privacy code , hypermeter (eps) value change do not effect the result.

initialize clients numbers

client_num = 10

load data

d = load_mnist(client_num)

lr = 0.01
fl_param = {
'output_size': 10, # number of units in output layer
'client_num': client_num, # number of clients
'C': 1,
'model': MLP, # model
'data': d, # dataset
'q': 0.1, # sampling rate
'tot_T': 100, # number of aggregation times
'E': 5, # number of local iterations
'batch_size': 64,
'lr': 0.01, # learning rate
'clip': 1, # clipping norm
'eps': 0.5, # privacy budget for each global communication

# 'eps': 50,                      # privacy budget for each global communication
'delta': 1e-5,                  # approximate differential privacy: (epsilon, delta)-DP
'device': device

}
import warnings
warnings.filterwarnings("ignore")
start_time = time.time()
fl_entity = FLServer(fl_param).to(device)
print('Currently performing FL with DP ---------------------------:')
print(datetime.datetime.now().strftime("%A, %d. %B %Y %I:%M%p"))
print(f'CPUs: {torch.get_num_threads()}, GPUs: {torch.cuda.device_count()} on {socket.gethostname()}.')
acc = []
loss = []
for t in range(fl_param['tot_T']):
fl_entity.set_lr(lr)
''' Update the local model according to weights update '''
loss_, acc_ = fl_entity.global_update()
loss += [loss_]
acc += [acc_]
print("Round = {:d}, loss={:.4f}, acc = {:.4f}".format(t+1, loss[-1], acc[-1]))
print(f'Total time: {time.time() - start_time :.2f} s.')

[alexandresablayrolles]

Can you please share more information, like the test accuracy you obtain at various levels of epsilon?

[Hafizamariaiqbal]

Can you please share more information, like the test accuracy you obtain at various epsilon levels?

Yes, first I execute code with epsilon value 0.5 with different noise added. In the next experiment, I did with the same parameter but change the epsilon value to 10, but no change after that I increase the value to 50 but the same results and epsilon in each round. No accuracy increases or decreases with the change of epsilon.

[zedoul]

I am experiencing the same issue, partly along with this reported issue. I am using neural network for tabular data.

I have experimented with different epsilon values, but the results remain relatively consistent. Interestingly, when using TensorFlow Privacy with the same model architecture, DP-optimizer, and hyperparameters, this problem does not arise.

[HuanyuZhang]

As an immediate step, could you print out the value of noise_multiplier (https://github.com/pytorch/opacus/blob/main/opacus/privacy_engine.py#L475)? That is the amount of noise added to the system. Sometimes when the noise difference is small, the model is robust to the noise change.