license: mit
pretty_name: Anonymization Before/After
size_categories:

• n<1K
  language:
• en
  task_categories:
• tabular-classification
• tabular-regression
• other
  tags:
• privacy
• anonymization
• k-anonymity
• synthetic-pii
• tabular
• teaching
• benchmark
• red-team
  configs:
• config_name: raw
  data_files: sample_raw.csv
• config_name: anonymized
  data_files: sample_anonymized.csv

Anonymization Before/After

A small paired tabular dataset showing the same records before and after
a 10-step anonymization pipeline. Useful as a teaching fixture for privacy
courses, a benchmark for anonymization toolkits, and a sanity-check input
for red-team / membership-inference experiments.

Important: the PII in sample_raw.csv is entirely synthetic.
Names follow the pattern Person_001, emails are [email protected],
phone numbers are 555-00XX, and "national IDs" are random 9-digit strings.
No real individual is represented. The dataset is published precisely so
researchers can demonstrate privacy techniques on PII-shaped data without
handling real PII.

At a glance

The 161 missing rows in anonymized were suppressed during k-anonymity
enforcement: every combination of (age band, region, gender) appears at
least 5 times in the released data.

What changed, column by column

After all transforms, columns in the anonymized split are sorted
alphabetically and rows are deterministically shuffled (seed=42).

How to use

from datasets import load_dataset

raw  = load_dataset("t22000t/anonymization-before-after", "raw",        split="train")
anon = load_dataset("t22000t/anonymization-before-after", "anonymized", split="train")

print(raw[0])
# {'full_name': 'Person_001', 'email': '[email protected]', ...,
#  'age': 51, 'region': 'Northeast', 'income': 40617.24, ...}

print(anon[0])
# {'age': '50 - 60', 'region': 'Other', 'income': 100899.98, ...}

Or plain pandas:

import pandas as pd
raw  = pd.read_csv("hf://datasets/t22000t/anonymization-before-after/sample_raw.csv")
anon = pd.read_csv("hf://datasets/t22000t/anonymization-before-after/sample_anonymized.csv")

Suggested uses

1. Teaching privacy concepts. The before/after diff makes k-anonymity,
   QI generalisation, and direct-ID suppression concrete in a single
   side-by-side view.
2. Benchmarking anonymization toolkits. Re-run a different tool on
   raw and compare its output to anonymized (or to your own
   adversarial baseline).
3. Red-team / membership-inference fixtures. Train an MIA classifier
   to separate raw from anonymized; the AUC quantifies how much
   identifying signal survived the pipeline.
4. Sanity-checking a synthetic-data generator. Use raw as the
   target distribution and compare against the generator's output on the
   same schema.

How it was produced

sample_raw.csv is generated by
scripts/create_sample_data.py
in the data-anonymization-toolkit
repo (seed=42, deterministic).

sample_anonymized.csv is the output of running the toolkit's full
10-step pipeline on raw with the config
config/example_simple_anonymization.yaml
(profile ml, k_target=5).

Reproduce locally:

git clone https://github.com/timothy22000/data-anonymization-toolkit
cd data-anonymization-toolkit
pip install -e .

python scripts/create_sample_data.py
python scripts/anonymize_data.py \
    --input data/sample.csv \
    --output data/anonymized.csv \
    --config config/example_simple_anonymization.yaml

Limitations and caveats

• Small. 500 rows is enough to demonstrate the pipeline behaviour but
  not enough for serious statistical evaluation of a generator.
• Synthetic PII. Names, emails, IDs are templated, not realistic.
  Adversarial fingerprinting attacks that exploit real-world string
  distributions (e.g. uncommon surnames) won't find anything interesting
  here. Use a realistic synthetic-PII library (Faker, mimesis) if that
  matters for your research.
• One profile, one config. Only the ml profile (k=5) is provided.
  The public profile (k=20, stronger generalisation) would suppress
  many more rows.
• The "Other" category dominates region. Because the toy generator
  uses 8 regions but top-N keeps only 6, two regions get collapsed to
  "Other" in the anonymized split. Real-world QI generalisation is
  usually more nuanced.

Personal and sensitive information

None. All PII-shaped fields are deterministically generated from row
indices. No real individual's data was used to create this dataset.

Citation

If you use this dataset in a paper or tutorial, please cite the toolkit
it came from:

@software{data_anonymization_toolkit,
  author = {Mun, Timothy},
  title  = {data-anonymization-toolkit: Config-driven anonymization,
            synthetic data, and red-team validation for tabular data},
  url    = {https://github.com/timothy22000/data-anonymization-toolkit},
  year   = {2026}
}

License

MIT - see the parent toolkit repo for the full license text.

Related

• 🔒 Privacy Lab Space - interactive anonymization + red-team demo (uses this dataset as the bundled sample)
• 🛡️ Synthetic Data Privacy Audit Space - upload real + synthetic CSVs, get a privacy verdict
• 📦 data-anonymization-toolkit - the underlying Python library