Netflix Prize Privacy Red-Team / Blue-Team Study

This project studies the privacy-utility tradeoff in the Netflix Prize dataset. It now has a concrete direction:

1. Use public IMDb user ratings as auxiliary information.
2. Match those ratings to Netflix Prize movie ids.
3. Run a probabilistic record-linkage attack against anonymous Netflix customer ids.
4. Create anonymized Netflix releases by removing/generalizing dates, coarsening ratings, adding noise, suppressing low-k facts, and removing rare movies.
5. Add ML red-team audits with sparse nearest-neighbor linkage and membership-inference classifiers.
6. Generate synthetic releases with SDV single-table synthesizers.
7. Compare privacy risk with k-anonymity, sampled linkage attacks, and ML attacks.
8. Compare utility with recommender baselines, including the package bias model and notebook PySpark ALS experiments.

The maintained package implementation lives in src/guardrails_sensitive_data. The active notebooks in notebooks/ are empirical studies and prototypes; the old notebooks in notebooks/old/ are historical exploration.

Ethical Scope

This repository is for a data privacy course/project. Run linkage attacks only on public IMDb profiles used for demonstration, and do not publish or contact candidate real-world identities. The purpose is to quantify privacy risk and evaluate defenses.

Data

The Netflix Prize data is not redistributed here. Place the official files in data/netflix/:

• combined_data_1.txt
• combined_data_2.txt
• combined_data_3.txt
• combined_data_4.txt
• movie_titles.csv or movie_titles.txt
• probe.txt for official holdout evaluation

This workspace already has the Netflix files locally, so normal commands can run without downloading anything.

Verify the local data:

python main.py verify-data --require-probe

If you have an authorized archive URL or local archive:

python main.py download-netflix --url "https://example.com/authorized/netflix.zip"
python main.py download-netflix --archive /path/to/netflix.zip

Environment

The recommended environment is Python 3.12. The ML notebooks use current compatible major versions of scikit-learn, SDV, and PySpark:

• scikit-learn>=1.9,<2.0
• sdv>=1.37,<2.0
• pyspark[sql]>=4.1,<4.2
• pyarrow>=15.0

PySpark 4.1 requires Java 17 or later with JAVA_HOME set. On macOS, for example:

brew install openjdk@17
export JAVA_HOME=$(/usr/libexec/java_home -v 17)

Using conda is the easiest path because environment.yml installs the package plus the notebook ML dependencies:

conda env create -f environment.yml
conda activate erdos_project_environment

Or with pip:

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ".[ml]"

For the lightweight CLI-only install, omit the ML extra:

python -m pip install -e .

The CLI works from the repository root even before installation:

python main.py --help

After installation, the same commands are available as:

netflix-privacy --help

Notebooks

Active notebooks:

• notebooks/01_synthetic_netflix_generator.ipynb uses SDV GaussianCopulaSynthesizer by default, with simple switches for CTGAN and TVAE.
• notebooks/02_ml_deanonymization_attacks.ipynb uses scikit-learn sparse vectors, NearestNeighbors, and LogisticRegression for red-team audits.
• notebooks/03_als_downstream_and_anonymization_feedback.ipynb uses PySpark MLlib ALS and item-factor diagnostics to guide targeted anonymization noise.
• notebooks/04_empirical_privacy_utility_study.ipynb is the cohesive report-style study that combines probabilistic linkage, anonymization, nearest-neighbor attacks, membership inference, synthetic data, downstream utility, plots, citations, and takeaways.
• final_pipeline.ipynb contains the final anonymization pipeline with the final privacy-utility curves.

Run notebooks from the repository root so relative imports and data paths line up:

jupyter lab

IMDb Ratings

The cached exploratory file notebooks/imdb_data.csv includes ratings for planktonrules and other public IMDb users. Use that for reproducible runs:

python main.py linkage-attack --user planktonrules

To fetch a public IMDb ratings page into a CSV, supply a user id or, preferably, the full /ratings URL:

python main.py scrape-imdb \
  --user planktonrules \
  --ratings-url "https://www.imdb.com/user/urXXXXXXXX/ratings" \
  --output reports/imdb_planktonrules.csv

IMDb markup changes often and may require login or JavaScript. Cached CSVs are the reliable research path.

Experiments

Run the probabilistic linkage attack:

python main.py linkage-attack --user planktonrules --top-n 50

Outputs:

• reports/linkage_planktonrules_matched_titles.csv
• reports/linkage_planktonrules_facts.csv
• reports/linkage_planktonrules_candidates.csv

Evaluate anonymization defenses:

python main.py privacy-eval --max-rows 1000000 --trials 300

Outputs:

• reports/privacy_k_anonymity_summary.csv
• reports/privacy_linkage_trials.csv
• reports/privacy_linkage_summary.csv

Compare downstream RMSE:

python main.py rmse-eval --max-rows 1000000

For the official probe holdout, use:

python main.py rmse-eval --holdout probe --max-train-rows 5000000 --max-probe-rows 100000

The default RMSE path uses a random holdout from a sampled subset, which is much faster and is enough to compare releases consistently.

Run a small end-to-end demo:

python main.py run-demo --max-rows 200000 --trials 50

For the full empirical report, open:

jupyter lab notebooks/04_empirical_privacy_utility_study.ipynb

The report notebook has runtime knobs at the top, including MAX_ROWS, PRIVACY_TRIALS, PROFILE_USERS, and SYNTH_TRAIN_ROWS.

Release Variants

The blue-team releases currently evaluated are:

• original movie + exact rating + month
• remove month
• generalize month to year
• coarsen rating into disliked / neutral / liked
• remove month and coarsen rating
• add bounded rating noise
• remove rare movies
• suppress low-k movie/rating/month facts
• movie only

movie_only is useful for privacy comparison but is skipped for RMSE because it does not release rating labels.

Method References

The notebooks cite the full reference list. The main methodological anchors are:

• Narayanan and Shmatikov, "Robust De-anonymization of Large Sparse Datasets" for the Netflix linkage threat model.
• Sweeney's k-anonymity model for fact suppression/generalization.
• scikit-learn nearest-neighbor and logistic-regression models for ML red-team audits.
• SDV, Gaussian Copula, CTGAN, and TVAE for plug-and-play synthetic tabular data generation.
• Spark MLlib ALS and matrix factorization for the stronger downstream recommender baseline.
• Membership-inference attacks for user-presence risk.

Tests

python -m unittest discover

The tests use tiny synthetic Netflix files and do not require the full dataset.