Federated Learning with Differential Privacy

This repository explores privacy-preserving federated learning with TensorFlow Federated (TFF). The main experiment trains an image classifier on federated EMNIST digits, applies user-level differential privacy to model aggregation, tracks the cumulative privacy budget, and compares several noise schedules in terms of privacy and model utility.

The repository also includes an earlier federated-learning prototype and the associated thesis report.

Repository contents

File	Purpose
final_(1).ipynb	Main EMNIST experiment with DP-FedAvg, privacy accounting, scheduled noise, validation, CSV logging, and comparative plots
TFF.ipynb	Exploratory TensorFlow Federated examples using EMNIST/MNIST, FedAvg, and initial differential-privacy experiments
Thesis_report.pdf	Thesis report describing the project and its findings

Main experiment

The main notebook uses the digits-only federated EMNIST dataset supplied by TensorFlow Federated. Each writer is treated as a client, preserving EMNIST's naturally federated structure.

The workflow includes:

• deterministic seeding for Python, NumPy, and TensorFlow;
• client-level training and validation splits;
• a multilayer perceptron with two 200-unit ReLU hidden layers;
• unweighted Federated Averaging with SGD client and server optimizers;
• differentially private model-update aggregation;
• Rényi differential privacy (RDP) accounting with dp-accounting;
• calibration against target privacy parameters;
• periodic validation and optional early stopping;
• per-round CSV logging of epsilon, accuracy, loss, and noise;
• privacy–utility analysis and comparison plots.

Default privacy configuration

The notebook defines the following initial experiment targets:

Parameter	Value
Target epsilon	3.0
Target delta	1e-5
Calibration rounds	100
Noise-to-clients ratio	0.01
Validation-client fraction	0.10
Evaluation interval	Every 5 rounds

The actual training helper accepts a configurable round limit and other evaluation/early-stopping settings.

Noise schedules

The main notebook implements and compares multiple noise schedules:

• flat: constant Gaussian noise;
• step5_low_to_high: five stages of increasing noise;
• step5_high_to_low: five stages of decreasing noise;
• cosine: smoothly decreasing noise using a cosine curve;
• linear_decreasing: linearly decreasing noise;
• exponential_decay: exponentially decreasing noise.

The included experiment cells directly run the flat and two five-step schedules. Other schedules can be selected through the same training function.

Reported endpoint comparison

The final notebook contains the following recorded endpoints for three runs:

Run	Test accuracy	Test loss	Rounds	Epsilon
Flat adaptive clipping	0.8126	0.5950	135	3.282
Step low-to-high	0.7721	0.7806	80	2.441
Step high-to-low	0.7850	0.7040	140	2.411

These values are embedded in the analysis notebook as saved experimental results; rerunning the complete experiment may produce different results due to software versions, runtime configuration, and stochastic training.

Running the notebooks

The notebooks are designed primarily for Google Colab. A fresh Colab runtime is recommended because TensorFlow Federated supports a limited range of TensorFlow and Python versions.

1. Open final_(1).ipynb in Google Colab.
2. Install the required packages in the first cell.
3. Restart the runtime if package installation requests it.
4. Run the setup, dataset, model, privacy-calibration, and training cells in order.
5. Update base_dir or BASE_DIR if Google Drive is not mounted at the paths used in the notebook.

A typical dependency set is:

!pip install tensorflow tensorflow-federated tensorflow-privacy dp-accounting \
    numpy pandas matplotlib

Use mutually compatible versions of TensorFlow, TensorFlow Federated, and TensorFlow Privacy. Installing their latest releases independently may result in dependency conflicts.

Outputs and analysis

Training runs log the following fields to CSV:

round, epsilon, train_acc, train_loss, val_acc, val_loss, sigma

The analysis cells can generate:

• validation accuracy and loss across rounds;
• training-versus-validation comparisons;
• epsilon and noise level across rounds;
• validation accuracy as a function of epsilon;
• accuracy gained per unit of privacy budget;
• area under the privacy–utility curve;
• three-run comparison tables and 3D plots.

By default, some cells read and write under /content/drive/MyDrive/tff_runs or /content. Change these paths before running locally or in a different Colab workspace.

Notes

• TFF.ipynb is an exploratory notebook and contains intermediate experiments; use final_(1).ipynb as the primary implementation.
• Differential-privacy guarantees depend on the sampling model, clipping, aggregation mechanism, and accountant assumptions. Recalculate epsilon after changing the number of clients, rounds, sampling rate, or noise schedule.
• The notebooks should be run sequentially in a clean runtime to avoid stale TensorFlow Federated state and incompatible model/process objects.