SNH

This repository contains the implementation of Spatial Neural Histograms [1] to answer range count queries on a geospatial dataset while preserving differential privacy. Neural networks are trained using JAX and are used to answer RCQs.

Instalation and requirements

Install conda environment and python targets:

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh

sh Miniconda3-latest-Linux-x86_64.sh

create a python target and source into python enivorment

conda create -n [name_of_enviornment] python=3.8

conda activate [name_of_enviornment]

install jax lib 0.1.71 with CUDA support (without GPU is also okay, ~10x performance penalty with 15 core CPU)

wget https://storage.googleapis.com/jax-releases/cuda111/jaxlib-0.1.71+cuda111-cp38-none-manylinux2010_x86_64.whl

pip install jaxlib-0.1.71+cuda111-cp38-none-manylinux2010_x86_64.whl

install other libraries

pip install jax==0.2.12 numpy pandas dm-haiku sklearn rtree

Running SNH

Running SNH can be done by calling python run.py. SNH configureations are set in run.py, through the python dictionary config. Specifically, the dictionary contains a key 'NAME'. When calling python run.py, the code creates the folder tests/config['NAME'], where the result of the experiment is written. Explanation of each of the configurations is available in file run.py.

SNH output

SNH trains config['no_models'] number of different models. The i-th model's training and testing statistic is written in the file tests/config['NAME']/i/out.txt

Example

The folder data contains two datasets: CABS_SFS.npy (from [2]) and gowalla_SF.npy (from [3]). We consider releasing CABS_SFS.npy with differential privacy while using gowalla_SF.npy as an auxiliary public dataset. Calling python run.py performs training and testing with this setting. For example, the result of the zero-th trained model will be written in tests/test_sf_cabs/0/out.txt. A sample output for that file is

Creating model for query size 0.2375±0.0375 % of query space

Preparing training data

Calculating training weights

initializing the model

training 100 Loss: 537.3705 mae: 26.526931762695312 rel. error: 0.11166811734437943 time : 8.425718907266855

200 Loss: 416.01135 mae: 23.483247756958008 rel. error: 0.10002146661281586 time : 13.059221280738711

300 Loss: 266.9216 mae: 21.457313537597656 rel. error: 0.09192200750112534 time : 17.616111068055034

400 Loss: 239.48528 mae: 20.5916690826416 rel. error: 0.0883064866065979 time : 22.172612854279578

500 Loss: 267.86072 mae: 20.23228645324707 rel. error: 0.0869283452630043 time : 26.740353140980005

600 Loss: 222.4536 mae: 19.401844024658203 rel. error: 0.08398811519145966 time : 31.336414461024106

700 Loss: 180.66728 mae: 19.385501861572266 rel. error: 0.08376768231391907 time : 35.92358886078

800 Loss: 222.31805 mae: 18.806110382080078 rel. error: 0.08121315389871597 time : 40.523882042616606

900 Loss: 140.94621 mae: 17.11213493347168 rel. error: 0.07491124421358109 time : 45.10136076621711

1000 Loss: 102.3436 mae: 17.200101852416992 rel. error: 0.07441579550504684 time : 49.67014822270721

1100 Loss: 213.91618 mae: 18.63749122619629 rel. error: 0.07999488711357117 time : 54.269442210905254

1200 Loss: 198.0504 mae: 18.650419235229492 rel. error: 0.08025692403316498 time : 58.86295662727207

1300 Loss: 91.117004 mae: 16.932586669921875 rel. error: 0.0732589140534401 time : 63.455598548054695

1400 Loss: 115.33469 mae: 17.047874450683594 rel. error: 0.07385049015283585 time : 68.03953194618225

1500 Loss: 188.22157 mae: 18.500423431396484 rel. error: 0.07881717383861542 time : 72.63241504132748

1600 Loss: 73.99296 mae: 17.00739097595215 rel. error: 0.07307520508766174 time : 77.20965842809528

1700 Loss: 119.67044 mae: 16.977487564086914 rel. error: 0.0728941336274147 time : 81.77187423035502

1800 Loss: 118.60206 mae: 16.82490348815918 rel. error: 0.07225397974252701 time : 86.3691204348579

1900 Loss: 74.22315 mae: 16.331043243408203 rel. error: 0.07074788212776184 time : 90.92316508665681

2000 Loss: 69.09277 mae: 16.455432891845703 rel. error: 0.0706830695271492 time : 95.49115407746285

2100 Loss: 169.36139 mae: 18.143535614013672 rel. error: 0.0776057243347168 time : 100.05525787267834

2200 Loss: 105.73973 mae: 16.141504287719727 rel. error: 0.06914427876472473 time : 104.67397096380591

2300 Loss: 112.43592 mae: 16.253826141357422 rel. error: 0.07009757310152054 time : 109.217095464468

2400 Loss: 136.64384 mae: 17.38785743713379 rel. error: 0.07459509372711182 time : 113.78008683677763

2500 Loss: 102.73479 mae: 16.36716079711914 rel. error: 0.07012166827917099 time : 118.3827364910394

2600 Loss: 66.67031 mae: 16.380956649780273 rel. error: 0.07002332806587219 time : 122.9673305992037

2700 Loss: 133.58382 mae: 17.387420654296875 rel. error: 0.07414235919713974 time : 127.65487424191087

2800 Loss: 140.5376 mae: 17.233060836791992 rel. error: 0.07426299154758453 time : 132.3707270808518

2900 Loss: 128.65556 mae: 16.50082015991211 rel. error: 0.07083127647638321 time : 137.0902198823169

3000 Loss: 103.58169 mae: 16.51291275024414 rel. error: 0.07087258249521255 time : 141.79305426962674

3100 Loss: 230.5178 mae: 17.801509857177734 rel. error: 0.07689779251813889 time : 146.44925714749843

3200 Loss: 50.446766 mae: 15.771015167236328 rel. error: 0.06772685796022415 time : 150.99709218740463

3300 Loss: 115.650055 mae: 16.614469528198242 rel. error: 0.0710296630859375 time : 155.58328399900347

3400 Loss: 144.73999 mae: 18.16712188720703 rel. error: 0.07756192237138748 time : 160.2031332720071

3500 Loss: 77.43564 mae: 16.369903564453125 rel. error: 0.07075335085391998 time : 164.77974649332464

3600 Loss: 44.42608 mae: 15.697988510131836 rel. error: 0.06757614761590958 time : 169.35493067558855

3700 Loss: 68.77671 mae: 16.144540786743164 rel. error: 0.06958482414484024 time : 173.91690846160054

3800 Loss: 125.346436 mae: 17.08391571044922 rel. error: 0.07286818325519562 time : 178.4751351652667

3900 Loss: 127.73309 mae: 17.198808670043945 rel. error: 0.07368797063827515 time : 183.0311270095408

4000 Loss: 132.26308 mae: 16.717212677001953 rel. error: 0.0713069960474968 time : 187.59608391765505

4100 Loss: 72.325836 mae: 15.670525550842285 rel. error: 0.06773883104324341 time : 192.1904318574816

4200 Loss: 105.53636 mae: 19.256271362304688 rel. error: 0.08139962702989578 time : 196.8051517298445

4300 Loss: 62.391106 mae: 15.962382316589355 rel. error: 0.0688060075044632 time : 201.39471916668117

4400 Loss: 138.56642 mae: 17.031509399414062 rel. error: 0.07295973598957062 time : 205.9722465183586

4500 Loss: 90.751175 mae: 16.03856086730957 rel. error: 0.06865857541561127 time : 210.55927829351276

4600 Loss: 82.98366 mae: 15.691493034362793 rel. error: 0.06760484725236893 time : 215.12370301876217

4700 Loss: 145.92674 mae: 17.186283111572266 rel. error: 0.07383159548044205 time : 219.68158520944417

4800 Loss: 90.90707 mae: 16.00307273864746 rel. error: 0.06884384155273438 time : 224.25662205833942

4900 Loss: 80.56631 mae: 15.907496452331543 rel. error: 0.0682239755988121 time : 228.83588003460318

5000 Loss: 32.029076 mae: 15.157988548278809 rel. error: 0.06545406579971313 time : 233.40768042951822

References

[1] Sepanta Zeighami, Ritesh Ahuja, Gabriel Ghinita, and Cyrus Shahabi, “A neural database for differentially private spatialrange queries, arXiv preprint: https://arxiv.org/abs/2108.01496

[2] Michal Piorkowski, Natasa Sarafijanovic-Djukic, and Matthias Grossglauser. 2009.CRAWDAD data set epfl/mobility (v. 2009-02-24)

[3] Eunjoon Cho, Seth A Myers, and Jure Leskovec. 2011. Friendship and mobility:user movement in location-based social networks. In Proceedings of the 17thACM SIGKDD international conference on Knowledge discovery and data mining.1082–1090