Skip to content

Latest commit

 

History

History
142 lines (88 loc) · 12.1 KB

README.md

File metadata and controls

142 lines (88 loc) · 12.1 KB
Raw

Differentially Private Top-k Selection

This repository provides an implementation of the FastJoint algorithm for differentially private top-k selection, as detailed in the paper:

Faster Differentially Private Top-k Selection: A Joint Exponential Mechanism with Pruning

Interface

The implementation can be found in FastTopk.py under the following function:

def fast_joint_sampling_dp_top_k(item_counts, k, epsilon, neighbor_type, failure_probability)

Arguments:

  • item_counts: A 1D numpy array representing the histogram (non-negative integer counts or scores for the items).
  • k: The number of items to select.
  • epsilon: The privacy parameter.
  • neighbor_type: The type of neighboring dataset. Currently supports "DP_Parameters.NeighborType.ADD_REMOVE" as defined in DP_Parameters.py.
  • failure_probability: The probability that the algorithm will return a sequence whose error exceeds the truncation threshold. The default value used in the experiments is $2^{-10}$.

Returns:
An array containing the indices of the top k items selected by the FastJoint algorithm.

Example Usage

import numpy as np

import DP_Parameters
from FastTopk import fast_joint_sampling_dp_top_k

# Create a histogram with counts from 20 to 2000 (1 to 100 multiplied by 20)
hist = np.arange(1, 101) * 20
k = 10
epsilon = 1
failure_probability = 2 ** (-10)

# Get the top k items using the FastJoint algorithm
top_k_items = fast_joint_sampling_dp_top_k(item_counts=hist, k=k, epsilon=epsilon,
                                           neighbor_type=DP_Parameters.NeighborType.ADD_REMOVE,
                                           failure_probability=failure_probability)

print(top_k_items)

Background

This section provides a brief overview of the problem and the algorithms used. For detailed information, please refer to the paper.

Given $d$ items indexed by $[d] = \lbrace 1, \ldots, d \rbrace$ and a histogram $\vec{h} = (\vec{h}[1], \ldots, \vec{h}[d]) \in \mathbb{N}_+^d$ representing their counts or scores, the algorithm returns an ordered sequence $\vec{s} = (\vec{s}[1], \ldots, \vec{s}[k])$ of $k$ distinct items from $[d]$, approximating the largest scores while ensuring differential privacy.

Neighboring Relation

Differential privacy relies on defining neighboring datasets. Two histograms $\vec{h}, \vec{h}' \in \mathbb{N}_+^d$ are considered neighbors if they differ by at most $1$ in at most one entry.

Algorithm

The FastJoint algorithm samples a sequence $\vec{s} = (\vec{s}[1], \ldots, \vec{s}[k])$ directly from the collection of $d^{\Theta(k)}$ possible length-$k$ sequences, using the exponential mechanism, based on the following loss function:

$$ \mathcal{E}rr(\vec{h}, \vec{s}) \doteq \max_{i \in [k]} \left( \vec{h}_{(i)} - \vec{h} \left[ \vec{s}[i] \right] \right), $$

where $\vec{h}_{(i)}$ is the $i^{th}$ largest entry in $\vec{h}$. The algorithm runs in $O\left( d + \frac{k^2}{\varepsilon} \cdot \ln d \right)$ time, where $\varepsilon$ is the privacy parameter.

Empirical Evaluation

This section provides instructions for reproducing the empirical results presented in the paper. It compares the FastJoint algorithm with the Joint, CDP Peel, and PNF Peel algorithms. The implementations of the later three algorithms can be found in the public dp_topk repository (as of May 2024). Please follow the instructions below to set up the environment, and download the necessary datasets, to replicate the results.

Code Integration

To integrate our modifications with the original dp_topk repository:

  1. Clone the dp_topk repository.
  2. Copy the files baseline_mechanisms.py, differential_privacy.py and joint.py from this repository.
  3. Place them into the dp_topk folder.

Datasets

To replicate our results, first download the datasets from the following sources:

  1. Goodreads Books: Goodreads-books
  2. Steam Video Games: Steam Video Games
  3. Tweets Dataset: Tweets Dataset
  4. Online News Popularity: UCI Online News Popularity
  5. MovieLens 25M Dataset: MovieLens 25M Dataset
  6. Amazon Product Data (2014): Amazon Grocery and Gourmet Food

Save the datasets as:

  • books.csv for the Goodreads dataset
  • games.csv for the Steam Video Games dataset
  • tweets.csv for the Tweets dataset
  • news.csv for the Online News Popularity dataset
  • movies.csv for the MovieLens dataset
  • foods.csv for the Amazon dataset

Place each file into a folder named datasets.

Note:
For the MovieLens 25M Dataset, the download will provide a zip file named ml-25m.zip. Extract its contents and locate the file ratings.csv. Rename this file to movies.csv and move it to the datasets folder.

Running the Experiments

To run the experiments, execute the following command:

python3 RunExp.py [num]

where [num] is an integer corresponding to a specific dataset:

Value Dataset
0 books
1 games
2 news
3 movies
4 tweets
5 food

The current code repeats each experiment 200 times. To speed up the code for obtaining preliminary results, you can reduce the number of trials by changing the initialization num_trials = 200 in RunExp.py to a smaller value.

For further details, refer to the file RunExp.py in the root directory.