De-identification and Best Practices

Last updated on 2026-08-27 | Edit this page

Overview

Questions

What does “de-identification” mean, and when is it required?
What techniques can I use to de-identify data?
How do I publish my data while protecting privacy?
What are the key best practices from this workshop?

Objectives

De-identify and prepare data for external transfer or publication
Apply de-identification techniques (removal, aggregation, generalization, perturbation)
Publish data in repositories while protecting sensitive information
Recognize course-wide best practices for data classification and handling

What Is De-identification?

De-identification removes (or obscures) information that could identify individuals, while preserving the research value.

Identifying Information

Identifiers that must be removed or obscured:

Category	Examples
Direct identifiers	Names, SSN, email, phone, address
Quasi-identifiers	Birth date, gender, ZIP code (in combination can re-identify)
Linkable identifiers	Medical record numbers, student ID numbers
Biometric data	Fingerprints, photographs, voice, DNA sequence
Sensitive attributes	Diagnosis, salary, grades (in combination can re-identify)

De-identification Techniques

1. Removal (Safest for Publication)

BASH

# Before (identifiable student data):
name,student_id,major,gpa
Alice Johnson,S12345678,Biology,3.85
Bob Smith,S87654321,Chemistry,3.72

# After (de-identified):
student_id,major,gpa
1,Biology,3.85
2,Chemistry,3.72

2. Aggregation (Reduces Detail)

BASH

# Before (identifiable):
student,exam_score
Alice,92
Bob,87
Carol,95

# After (aggregated):
major,mean_score,std_dev,n
Biology,89.3,3.5,15
Chemistry,87.1,2.8,18

3. Generalization (Reduces Precision)

BASH

# Before (specific, re-identifiable):
birth_date,diagnosis
1989-03-15,Type 2 Diabetes

# After (generalized):
birth_year,diagnosis
1989,Type 2 Diabetes

4. Perturbation (Adds Noise)

BASH

# Before:
name,height_cm,weight_kg
Alice,165,58

# After (add random noise):
height_cm,weight_kg
166±2,59±3

De-identification Workflow

START: Identifiable research data (e.g., student survey)
   ↓
1. Identify what's identifiable
   (Names? IDs? Dates? Locations?)
   ↓
2. Choose de-identification method
   (Remove, aggregate, generalize, perturb?)
   ↓
3. Apply de-identification
   (Create new dataset from original)
   ↓
4. Verify no re-identification is possible
   (Could someone piece it back together?)
   ↓
5. Keep original (encrypted) separate from de-identified
   ↓
6. Share de-identified version

Publication Workflow

To publish data based on Restricted sources: (1) de-identify using the techniques above, (2) clean data and add documentation (README, metadata, codebook, license), (3) choose a repository (Zenodo, OSF, NCBI, etc.), (4) upload and receive a DOI, (5) cite the DOI in your paper. Always keep the original identifiable data encrypted and separate from the de-identified version. Store any mapping files (original ID to anonymous ID) in your gocryptfs encrypted container.

Course Summary

You’ve completed Workshop 14: Data Classification and Handling. You can now:

Classify research data using Pomona’s three-tier system (PUBLIC/PROPRIETARY/RESTRICTED)
Implement security controls appropriate to each tier
Store and encrypt Restricted data using gocryptfs (AES-256-GCM)
Share data with collaborators while maintaining compliance
De-identify sensitive data for publication
Recognize and report data security incidents

Next Steps:

Inventory and classify your lab’s data: Use the classification worksheets
Set up encryption: Create gocryptfs containers for Restricted data
Document access controls: Create access authorization lists
Plan retention and deletion: Know when data can be discarded

For questions: - Data classification / Encryption / Storage: its-hpc@pomona.edu - Export control: orsp@pomona.edu - IRB/FERPA questions: orsp@pomona.edu

Additional Resources

Pomona Documents: - Pomona Information Security Policy - HPC Usage Policy - Sagehen HPC User Guide - Data Classification Guide

External Standards: - NIST SP 800-171 . Safeguarding CUI - FERPA Overview . Student Privacy Law - EAR . Export Controls - HIPAA . Health Information Privacy

Tools: - gocryptfs Documentation - rclone . Cloud storage tool

Challenge

Challenge 12.1: De-identification Assessment

You have health survey data (not HIPAA-covered, but sensitive):

name,email,phone,birth_date,zip_code,medical_condition,medication
Jane Smith,j.smith@email.com,555-0123,1985-04-15,91010,Type 2 Diabetes,Metformin
John Doe,j.doe@email.com,555-0124,1990-07-22,91010,Hypertension,Lisinopril
Jane Loo,j.loo@email.com,555-0125,1988-11-30,91010,Type 2 Diabetes,Metformin

Tasks: 1. Identify all re-identification risks (direct, quasi, sensitive) 2. Propose de-identification strategy 3. Create de-identified version 4. Assess residual re-identification risk

Show me the solution

Risks: Names (direct), emails (direct), phone (direct), birth_date (quasi), zip_code (quasi, only 3 records for 91010), medical_condition (sensitive), medication (sensitive)

De-identification: - Remove: names, emails, phones, exact birth dates - Generalize: birth_date → year only; zip_code → state - Aggregate: medical_condition + medication together

Result:

birth_year,state,condition_medication
1985,CA,Diabetes/Metformin
1990,CA,Hypertension/Lisinopril
1988,CA,Diabetes/Metformin

Residual risk: Low (no direct identifiers, quasi-identifiers generalized, aggregated condition/med)

Key Points

De-identification removes identifiers while preserving research value
Four techniques: removal, aggregation, generalization, perturbation
Keep mapping files encrypted and separate from de-identified data
Publish data with DOI for reproducibility and impact
Data use agreements formalize access and protection obligations
Proper data classification protects your research, your community, and Pomona College