SAMBA: a generic framework for Secure federAted Multi-armed BAndits

Radu Ciucanu, Pascal Lafourcade, Gael Marcadet, and Marta Soare.

Journal of Artificial Intelligence Research (JAIR)

If you use our code, please cite

@article{CLMS22,
  author    = {Ciucanu, R. and Lafourcade, P. and Marcadet, G. and Soare, M.},
  title     = {{SAMBA: A Generic Framework for Secure Federated Multi-Armed Bandits}},
  journal   = {Journal of Artificial Intelligence Research (JAIR)},
  volume    = {73},
  pages     = {737--765},
  year      = {2022}
}

Installation Guide

This section describes the way to install Samba in a local environment.

Requirements

To run Samba, we need to have some library already installed on your system. To install them,

apt-get install build-essential git python3 python3-pip python3-venv -y

Installation of Samba

Samba needs some libraries with a specific version in order to guarantee a functional program over the time.

The first thing to do is to download the source from git by typing:

git clone https://github.com/gamarcad/paper-samba-code.git
cd paper-samba-code

To limit the impact of these libraries on the installed system, we provide a Python virtual environment, allowing to install all libraries we need in a specific location out of your python system configuration.

To install the virtual environment, type the command

./setup_virtual_env.sh

This script creates and initializes the virtual environment with all desired libraries installed. Once executed, a new directory called 'samba' will be created, containing python binaries. The virtual environment can be managed with these commands to enter in your terminal:

• source samba/bin/activate opens an instance in the virtual environment. Once executed, you will notice (samba) in front of your prompt (e.g., (samba) user@machine $), meaning that you are in the virtual environment. Note that all your commands installed on your system will be usable as usual.

• deactivate exits the virtual environment. Works only when you are in the virtual environment, this command will only affect the instance in the virtual environment, meaning that the current path will not change.

Troubleshooting

When an error occurs during the first steps of the installation, please start by updating your system, by doing sudo apt-get update -y && sudo apt-get upgrade -y.

If an error occurs during cryptography library installation, causing by a missingsetuptools_rust module, you need to update the pip module inside the environment by doing samba/bin/python3 -m pip install --upgrade pip command. The solution is adapted from the issue claimed at pyca/cryptography#5753

Benchmark Execution Guide

This section describes how to reproduce the plots used in the article.

Once the virtual environment is installed, you will be able to run the benchmark Enter the virtual environment and run the following commands in the presented order:

• sudo python3 bench.py -h This command displays details of each parameter needed by the benchmark program. You can complete each required option, or to execute the next command which is basically a shell script which calls the bench.py python script with appropriate arguments. Some files must be indicated to the program:

  • The --configs <filename> option indicates the file containing all pairs of numbers of arms K and budget N to consider, in a CSV file format.

  • The --data <filename> option indicates the file containing arms probabilities in a textual format, one probability by line.

  • Finally, the --parameters <filename> option indicates the file containing parameters in a CSV format, used to feed algorithms that requires some parameters.

  • To decrease the needed execution time, bench.py supports the multi-processing. By default, it uses the number of core available on the system. You can decide the number of core to use with --cpu <nb_core> option.

  • In case you want to focus on some algorithm, you can indicate the --algorithms <IDalgo1> <IDalgo2> ... where supported algorithms, detailed here by name and ID are:

Algorithm	ID
Pursuit	pursuit
Softmax - Boltzmann	boltzmann
epsilon-greedy	e-greedy
epsilon-greedy decreasing	e-decreasing-greedy
Thompson Sampling	thompson-sampling
UCB	ucb

• sudo ./launch_benchmarks.sh This command executes the benchmark with pre-filled parameters. Note the --run option, which indicates to the program to run the benchmarks. Benchmarks may take a long time depending on configurations and data. To ensure that the given configuration is well stated, by default the benchmarks program displays the configuration based on the inputs. When you are sure about the configuration, then indicate the --run option.

  Additionally, this scripts calls the plotting program which generates plots, under the output directory. Hence, no more action is necessary to get the plots.

Only Plotting Guide

The benchmark take a long time as the number of configuration to be tested is high. In this repository, we provide our pre-computed results located under the precomputed_results folder, in a JSON format.

You can generate plots we used in the paper by executing the sudo ./only_plots.sh shell script which displays the plots under precomputed_results/plots folder. There are 4 plots:

• rewards.pdf plots the cumulative rewards with respect to the budget N.
• scalability_N.pdf plots the execution time while varying the budget N with K = 100.
• scalability_K.pdf plots the execution time while varying the number of arms K with N = 100000.
• zoom.pdf plots the execution time for each entity involved in the process with N = 100000 and respectively K = 10 and K = 100.

Detailed Data Files Format and Generation

We detailed in this section the format of each file given to the benchmark python script.

Data Files

Data files, located under the data folder, contain the list of probabilities used in the multi-armed bandits algorithms. Each probability is placed alone on a single line. At the beginning of the data file is located the number of probabilities present in the concerned data file.

In our experiments, we used MovieLens and Jester datasets, that we preprocessed in order to adapt them for multi-armed bandits.

Dataset	Location
MovieLens	http://files.grouplens.org/datasets/movielens/ml-100k/
Jester	http://eigentaste.berkeley.edu/dataset/

MovieLens

The MovieLens dataset is composed of 100 movies rated by 943 number of users, where each rating is between 1 and 5. To adapt MovieLens for multi-armed bandits, we compute for each movie his the number of ratings greater or equals than a given threshold, fixed at 4 , divided by the number of ratings of the concerned movie.

Jester

The Jester dataset contains 100 jokes rated by 24983 users, where each rating is between -10 for a very bad joke and 10 for a very good one. Like the MovieLens preprocessing, for each joke, we divide the number of rating greater or equals of a given threshold, fixed at 4.5 , by the number of ratings for the concerned joke.

To pre-processed these files, we used the preprocessing.py scripts. You do not need to execute this script to launch the experiment as long as data files exists. In case you want to launch experiments on data generated with a different threshold, then modify the thresholds variables in the preprocesing.py python script and execute it with python3 preprocessing.py without arguments.

Parameter Files

Parameter files located under parameters folder, contains parameters optimized for datasets we used in our experiments. Each parameter file is a CSV file, containing the header row tau, beta, epsilon columns and a single data row containing parameter in order defined in the header.

We used these parameters thanks to the parameters_suggestion.py script which, for a given dataset, suggests the set of parameters which returns the highest cumulative reward.

Config Files

In the paper, we show experiment both on the execution time and the cumulative reward. Depending on of the desired experiments, the list of possible combinations of number of arms K and budget N to test will not be the same. It is why we create a configuration file, located under configs folder, in a CSV format, with a header row N, K and data rows, one for each combination to test.