Differentially Private Heatmaps

Authors: Badih Ghazi, Junfeng He, Kai Kohlhoff, Ravi Kumar, Pasin Manurangsi, Vidhya Navalpakkam, Nachiappan Valliappan

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We test our algorithm on both real-world location datasets and synthetic datasets. The results demonstrate its practicality even for moderate values of ε [0.5, 5] and a number of users equal to 200. Furthermore, we compare our algorithm with simple baselines; under popular metrics for heatmaps, our results demonstrate significant improvements on these regimes of parameters.
Researcher Affiliation Industry Google Research, Mountain View, CA {badihghazi, ravi.k53}@gmail.com {junfenghe,kohlhoff,pasin,vidhyan,nac}@google.com
Pseudocode Yes The paper includes 'Algorithm 1: DPSPARSEEMDAGG' and 'Algorithm 2: RECONSTRUCT' in pseudocode format.
Open Source Code No The paper does not provide any explicit statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes We use two datasets available at snap.stanford.edu to generate the input distribution for users. The first dataset, called GOWALLA... Available at http://snap.stanford.edu/data/loc-Gowalla.html. The second dataset, called BRIGHTKITE... Available at http://snap.stanford.edu/data/loc-Brightkite.html.
Dataset Splits No The paper does not explicitly provide details about training, validation, and test dataset splits (e.g., percentages, sample counts, or citations to predefined splits) for the heatmaps. It describes sampling users for runs but not splitting the grid data into separate train/validation/test sets.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU, GPU, or TPU models, or memory amounts) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers (e.g., 'Python 3.8, PyTorch 1.9') used for replicating the experiment. While it mentions 'TensorFlow Privacy' and 'PyTorch Differential Privacy' in related work, it does not state these were used for its own implementation with version numbers.
Experiment Setup Yes As for our parameters, we use the decay rate γ = 1/√2, which is obtained from minimizing the second error term in the proof of Theorem 3.1 as ℓ 5. We use w = 20 in our experiments, which turns out to work well already for datasets we consider.