yangfan-jiang / federated-learning-with-differential-privacy Goto Github PK

fl_param = {
'output_size': 10,
'client_num': client_num,

'model': MNIST_CNN,
'data': d,
'lr': lr,
'E': 200,
'C': 1,
'eps': 1.0,
'delta': 1e-5,
'q': 0.01,
'clip': 0.1,
'tot_T': 30,
'batch_size': 128,
'device': device

}

if I want to preset (epsilon, delta) pair as privacy guarantee, what should I do?
just modify FLClient.update, not change the sensivity, but modify new_param[name] += add guassian noise defined in gaussian_noise(), will it work?
I want to use the code of
def gaussian_noise()
Waht should I do? Recalculate the sensitivity?
Thanks!!!

I meet some trouble when I apply differential privacy into FL in the way of CDP.Specifically，I add noise during aggregated stage in server, and then broadcast to all client. However, the finally accuracy result is low……
I'd be appreciated it if you could do me some favor.

请问作者，我想对比一下不加噪声和加噪声最后的结果。如何配置能使得这个联邦学习不添加噪声

why your sensitivity = 2 * self.lr * self.clip / self.data_size + (self.E - 1) * 2 * self.lr * self.clip? in the paper ,the sensitivity = 2C*pi/m or 2C/m.

Nice job!
Where can I reference the way that you compute the sensitivity
sensitivity = 2 * self.lr * self.clip / self.data_size + (self.E - 1) * 2 * self.lr * self.clip

By the way, are there any other implementations of dp algo except the Gaussian?

Hello : )

I used global_update_grad() because I tried to playback the result in the original paper.
I'v already added epsilon and delta in class FLSever and fl_param, but it still ran an error like this
Did I make a mistake or miss something?
Thank you !