Skip to content

Code for "ObscureNet: Learning Attribute-invariant Latent Representation for Anonymizing Sensor Data", ACM/IEEE IoTDI 2021

License

Notifications You must be signed in to change notification settings

sustainable-computing/ObscureNet

Repository files navigation

ObscureNet

This repository contains the implementation of ObscureNet and the baselines proposed in our IoTDI'21 paper entitled "ObscureNet: Learning Attribute-invariant Latent Representationfor Anonymizing Sensor Data".

Each directory is named after a privacy-preserving method described in the paper.

Datasets

The 2 Human Activity Recognition (HAR) datasets used to evaluate different methods are MotionSense and MobiAct. You can download them from the following websites and use the provided converter (dataset_builder.py) to preprocess the data and turn it into the format that our code expects:

To reproduce the results of our paper, use the CSV file dataset_subjects, which is provided in this repo, instead of the original one that comes with the MobiAct dataset.

Dependencies

Package Version
Python3 3.6.9
Tensorflow 1.14.0
PyTorch 1.4.0
Keras 2.3.1

How to cite ObscureNet

Omid Hajihassani, Omid Ardakanian and Hamzeh Khazaei. 2021. ObscureNet: Learning Attribute-invariant Latent Representation for Anonymizing Sensor Data, In Proceedings of the 6th ACM/IEEE Conference on Internet of Things Design and Implementation (IoTDI).

About

Code for "ObscureNet: Learning Attribute-invariant Latent Representation for Anonymizing Sensor Data", ACM/IEEE IoTDI 2021

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •  

Languages