lainisourgod/hybrid-approach-to-ppfl

This is an implementation for paper "A Hybrid Approach to Privacy Preserving Federated Learning" (https://arxiv.org/pdf/1812.03224.pdf)

PoC Hybrid Approach

Project structure

• main.py -- script for MNIST experiment (either Linear or CNN model). Defines dataset and dataloaders for trainer
• rnn_main.py -- script for RNN experiment
• trainer.py -- Trainer class that rules training and encryption stuff
• distro.py -- Party and Server classes that knows about model and about encryption and do it
• distro_paillier/ -- sources of Threshold Paillier Homomorphic Encryption. this library is used. I've done minor changes to decrypt function for it to work correctly with float numbers.
• config.py -- containes singleton config object used by all modules.

How to run

1. Install system dependencies for crypto stuff: sudo apt install libgmp-dev libmpfr-dev libmpc-dev.
2. Install python dependencies: pip install -r requirements.txt.
3. Run MNIST experiment python main.py.
4. Run RNN experiment python rnn_main.py.

hybrid-approach-to-ppfl's Issues

aggregate_params err

(hybrid) ubuntu@ubuntu-SYS-7049GP-TRTA-AI034:Workspace/hybrid-approach-to-ppfl$ python main.py
Bit-length of primes: 128
Starting generation of p, q and N...
Generation successful.
Number of insuccesfull trials small prime test: 1330
Number of insuccesfull trials biprime test: 16
Elapsed time: 1 seconds.
Traceback (most recent call last):
File "Workspace/hybrid-approach-to-ppfl/main.py", line 84, in
trainer.fit()
File "Workspace/hybrid-approach-to-ppfl/train.py", line 72, in fit
self.fit_on_batch()
File "Workspace/hybrid-approach-to-ppfl/train.py", line 97, in fit_on_batch
aggregate: np.ndarray = self.server.aggregate_params(encrypted_models)
File "Workspace/hybrid-approach-to-ppfl/distro.py", line 72, in aggregate_params
return np.mean(gradients_of_parties, axis=0)
File "<array_function internals>", line 200, in mean
File "/home/ubuntu/anaconda3/envs/hybrid/lib/python3.9/site-packages/numpy/core/fromnumeric.py", line 3464, in mean
return _methods._mean(a, axis=axis, dtype=dtype,
File "/home/ubuntu/anaconda3/envs/hybrid/lib/python3.9/site-packages/numpy/core/_methods.py", line 165, in _mean
arr = asanyarray(a)
ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 2 dimensions. The detected shape was (10, 2) + inhomogeneous part.

keyword-only arguments missing

Error when running main.py

Bit-length of primes:  128
Starting generation of p, q and N...
Generation successful.
Number of insuccesfull trials small prime test:  323
Number of insuccesfull trials biprime test:  4
Elapsed time:  0 seconds.
Traceback (most recent call last):
  File "main.py", line 77, in <module>
    trainer = Trainer(
  File "/home/student/Desktop/hybrid-approach-to-ppfl-master/train.py", line 40, in __init__
    self.configure_system()
  File "/home/student/Desktop/hybrid-approach-to-ppfl-master/train.py", line 50, in configure_system
    self.parties = [
  File "/home/student/Desktop/hybrid-approach-to-ppfl-master/train.py", line 51, in <listcomp>
    Party(model=self.model, pubkey=self.server.pubkey)
  File "/home/student/Desktop/hybrid-approach-to-ppfl-master/distro.py", line 109, in __init__
    self.randomiser = dp.mechanisms.Gaussian().set_epsilon_delta(1, 1).set_sensitivity(0.1)
TypeError: __init__() missing 3 required keyword-only arguments: 'epsilon', 'delta', and 'sensitivity'