Table of Contents
This framework is an end-to-end implementation of the protocol proposed in of RoFL: Attestable Robustness for Secure Federated Learning. The protocol relies on combining secure aggregation with commitments together with zero-knowledge proofs to proof constraints on client updates. One constraint that we show to be effective against some types of backdoor attack is the bounding of update norms. We evaluate this constraint using the federated learning analysis framework, which can be used to perform experiments to analyse the effectiveness of various federated learning backdoor attacks and defenses.
The current implementation of RoFL is an academic proof-of-concept prototype. This prototype is designed to focus on evaluating the overheads of zero-knowledge proofs on client updates on top of secure aggregation. The current prototype is not meant to be directly used for applications in productions.
This repository is structured as follows.
- RoFL_Service: This directory contains the code for the federated learning server and client, written in Rust.
- RoFL_Crypto: This directory contains the cryptographic library to generate and verify the zero-knowledge proof constraints used in RoFL.
- RoFL_Train_Client: This directory contains a python service to handle the training and inference operations for the machine learning model. The RoFL_Service interfaces with the RoFL_Train_Client using gRPC. This component acts as a wrapper around the FL Analysis framework that is used for machine learning training.
-
ansible: The Ansible setup used for the evaluation of the framework. For more information on how to use this, See ansible/README.md.
-
plots: This directory contains code used to generate plots for the paper.
The end-to-end setup consists of two components. First the secure federated learning with constraints implementation performs the communication between the server and the clients. On the client- and server-side, this component offloads the machine learning operations to a python training and evaluation service.
Follow these steps to run the implementation on your local machine.
Python 3.7 & Rust version:
Min: 1.52.0-nightly
Both the secure FL component and the training service are installed separately.
- Clone this repository
git clone [email protected]:pps-lab/rofl-project-code.git
- Install Cargo/Rust
curl https://sh.rustup.rs -sSf | sh -s -- -y
- Switch to nightly. Note: As of now, only a specific nightly version is supported due to a deprecated feature that a dependency is using.
rustup override set nightly-2021-05-11
- Select specific versions for some dependencies (due to these dependencies breaking some nightly features)
cargo update -p proc-macro2 --precise 1.0.28
cargo update -p packed_simd_2 --precise 0.3.4
cargo build
- Install the requirements for the trainservice wrapper (in
rofl_project_code
)
cd rofl_train_client
pip install -r requirements.txt
- Download the analysis framework
cd ../../ # go up to workspace directory
git clone [email protected]:pps-lab/fl-analysis.git
- Install the requirements for the analysis framework
cd fl-analysis
pipenv install
The framework can be used in two ways.
We provide a setup in Ansible to easily deploy and evaluate the framework on multiple servers on AWS. See ansible/README.md for instructions on how to use this Ansible setup.
To run the setup manually, several components must be run separately. The individual components must be started in this order. The example shown are for a basic local configuration with four clients with L8-norm (infinity) range proof verification. The implementation of RoFL uses the analysis framework for model training and evaluation. In the following, we assume the following directory structure:
Top-level directory (e.g., workspace):
rofl-project-code
(this repository)fl-analysis
(the analysis framework)
Each component must be run in a separate terminal window.
In rofl-project-code
, run the server:
./target/debug/flserver
First, navigate to the analysis framework directory and enter the pipenv:
cd ../fl-analysis
pipenv shell
Then, navigate back to the python directory in the implementation directory:
cd ../rofl-project-code/rofl_train_client
From the rofl_train_client
directory, run the python service.
cd rofl_train_client
PYTHONPATH=$(pwd) python trainservice/service.py
In the rofl-project-code
directory, run the client executable.
cd ../
./target/debug/flclients -n 4 -r 50016
After running the client, training has started. In addition, the observer component can be used to evaluate the model accuracy on the server-side. To do so, first, navigate to the analysis framework directory and enter the pipenv:
cd ../fl-analysis
pipenv shell
Then, navigate back to the python directory in the implementation directory:
cd ../rofl-project-code/rofl_train_client
Set the PYTHONPATH to include the current directory and run
PYTHONPATH=$(pwd) python trainservice/observer.py
The observer will connect to the FL server and receive the global model for each round.
The implementation outputs time and bandwidth measurements in several files.
The benchmark files for both the server and the clients can be found in the benchlog folder.
#### Format of the server log
t1--t2--t3--t4
t1: round starts
t2: round aggregation done
t3: round param extraction done
t4: verification completes
Format of a benchmark log line:
<Round ID>, <t2 - t1>, <t3 - t2>, <t4 - t3>, <total duration>
#### Format of the client log
t1--t2--t3--t4--t5
t1: model meta received
t2: model completely received
t3: local model training done
t4: model update encryption + proofs completed
t5: model sent to server
Format of a benchmark log line:
<Round ID>, <t2 - t1>, <t3 - t2>, <t4 - t3>, <t5 - t4>, <total duration>, <bytes received>, <bytes sent>
This project's code is distributed under the MIT License. See LICENSE
for more information.
Project Links: https://github.com/pps-lab/rofl-project-code and https://pps-lab.com/research/ml-sec/