Anonymization of High-Dimensional Health Data

using in-memory applications based on SAP HANA for exact anonymization through suppression, generalization, perturbation and compartmentation.

This repository holds the source code of the MA thesis Anonymization of High-Dimensional Health Data as practicality proof-of-concept implementation. Besides implementing the exact identification of quasi-identifier which is W[2]-complete, different treatment methods are implemented. This repo also contains the necessary functionality the test, analyze and evaluate the anonymization results.

Repo Structure

.
├── cache
│   ├── .keep
├── images
│   ├── .keep
├── l_diversity.py
├── log
│   ├── .keep
├── output
│   ├── .keep
├── credentials.json
├── evaluation.py
├── quick_eval_score_dataset.py
├── README.md
├── requirements.txt
├── scores.txt
└── treatment_eval.sh
├── analyzer.py
├── anonymizer-api.py
├── anonymizer-job.py
├── anonymizer.py

• The cache directory holds the serialized mpmUCCs combinations (2nd class identifiers). When using this caching functionality, the identification step can be skipped reducing the processing time for multiple identification iterations on the same dataset.
• The images directory contains visualizations of the processing if enabled. This can be the graphical representation of the 2nd class identifier as Graph or the process of creating compartments as maximal cliques in the inverted graph.
• In the output directory are various files which are generated during the anonymization, including the sanitized dataset, statistics and much more.
• The log directory holds log files of the anonymization procedure.

Dependencies & Setup

Overall

sudo apt-get update
sudo apt-get install -y python3 python3-pip python3-numpy python3-scipy python3-matplotlib redis-server

Use the newest python version (at least 3.6.2), since in the previous versions there is a rare bug in the multiprocessing engine leading to a deadlock.

sudo add-apt-repository ppa:jonathonf/python-3.6
sudo apt-get update
sudo apt-get install python3.6

API service specific

For the API service, kindly provide a local redis-server by installing:

wget http://download.redis.io/redis-stable.tar.gz
tar xvzf redis-stable.tar.gz
cd redis-stable
make

Keep in mind to adjust the settings in /etc/redis/redis.conf regarding to your needs. Afterwards restart the service through sudo service redis-server restart.

One may start the API service manually via python3 anonymizer-api.py or use phusion passenger for running python apps in combination with a apache2 webserver for a production ready setup.

Have a SAP HANA up and running including valid configuration, extensions and user management.

Configuration

Make sure that an SAP HANA instance is running as well as a local redis server. Modify the credentials for HANA and redis in the credentials.json file.

For connecting with python to the HANA instance pyhdb or sqlalchemy is necessary. The former can be found at github.com/SAP/pyhdb or simply through pip3 install pyhdb.

For the lib sqlalchemy_hana in /local/lib/python3.5/dist-packages/sqlalchemy_hana/dialect.py all instances of unicode must be replaced with str.

API Design

Executing an http post request triggers the action of the synchrous process by sampling from the dataset and checking its anonymity once again. The user will receive the resource details including the table, where the data view is available one.

HTTP POST http://localhost:5000/jobs

{
  "nrows":1000,
  "columns":["Age","Gender","Surname"],
  "partner_id":0
}

Following a RESTful service API, a GET requests details for a specific resource, DELETE removes the resource and PUT may be used to update settings in the resource.

For instances by using:

• HTTP GET http://localhost:5000/jobs/1
• HTTP PUT http://localhost:5000/jobs/1
• HTTP DELETE http://localhost:5000/jobs/1

Anonymizer.py

The anonymizer.py file can be called directory, offering the following arguments:

Argument (short)	Argument (long)	Description	Default	Example
-ncol	--columns	Amount of columns to be considered	12	78
-exact	--exact	Determining 2nd class identifier exact instead of heuristic	no	{yes,no}
-treat	--treatment	Apply treatment?	yes	{yes,no}
-evaltreat	--eval_treatment	Shall treatment be evaluated?	yes	{yes,no}
-treatment	--overwrite_treatment	Overwrite automated treatment which predefined one	no	compartmentation
-analyze	--analyze_results	Shall quasi identifier be analyzed?	no	{yes,no}
-score	--data_value_score	Shall data value be evaluated?	no	{yes,no}
-size	--data_size	Shall data size be logged?	no	{yes,no}
-visualize	--visualize	Shall visualize be created?	no	{yes,no}
-cache	--cache	Shall the mpmUCCs / 2nd class identifier be cached?	yes	{yes,no}