Oblivious (Un)Learning of Extremely Randomized Trees using TFHE

Dependencies

You need to install Rust and Cargo to use tfhe-rs and RevoLUT.

First, install the needed Rust toolchain:

rustup toolchain install nightly

Then, you can either:

1. Manually specify the toolchain to use in each of the cargo commands: For example:

cargo +nightly build
cargo +nightly run

2. Or override the toolchain to use for the current project:

rustup override set nightly

Cargo will use the nightly toolchain.

cargo build

Usage

The program supports three execution modes: standard, hybrid and oblivious. The three execution modes supported by this program are as follows:

• Standard Mode:
  In this mode, the forest of extremely randomized trees (ERTs) is fully trained on unencrypted (clear) data. After training, the best-performing model is selected (optionally using multiple trials), exported (without count leaves), and is retrained on encrypted data. This mode is useful for benchmarking and verifying correctness.

• Hybrid Mode:
  The dataset is split into two subsets, D_0 and D_1. The ERTs are first trained on D_0 in the clear using the Gini-index criterion, and counts for classes at each leaf are recorded (exported with the model). Then, additional training (or updating) can be done using encrypted data from D_1.

• Oblivious Mode (Oblivious Training/Unlearning):
  This mode allows you to perform oblivious training or unlearning of the forest on encrypted data from a provided CSV file. The same function is used for both operations; labels are encoded differently to indicate whether you are training (add counts) or unlearning (remove counts).

Standard Mode

The standard mode trains a forest in clear, selects the best model, exports it without counts, then retrains with encrypted data to demonstrate that training with encrypted data does not degrade accuracy.

Example:

cargo run --release -- --mode standard --dataset iris --num-trees 8 --depth 4  --split 0.8 --best-model-trials 10 --verbose

Options:

• --dataset or -d: Dataset name (iris, wine, cancer). Default: iris
• --num-trees or -n: Number of trees (comma-separated values, e.g., "8,16"). Default: 8,16
• --depth: Tree depth. Default: 4
• --split: Train/test split percentage. Default: 0.8
• --trials or -t: Number of repetitions/trials. Default: 10
• --best-model-trials: Number of trials to find the best model. Default: 1
• --output: Output directory. Default: ./export/
• --verbose or -v: Verbose mode
• --mode: Execution mode (standard or hybrid). Default: standard

Hybrid Mode

The hybrid mode, is more "practical", it splits the dataset into D_0 and D_1, trains ERTs in clear on D_0 using Gini-index, exports the forest with counts, then continues training on encrypted D_1.

Example:

cargo run --release -- --mode hybrid --dataset iris --num-trees 8 --depth 4  --split 0.3 --verbose

Options:

• Same as standard mode, with the following differences:
• --split: Used for D_0/D_1 split (percentage for D_0, remainder is D_1). The train/test split within D_0 is fixed at 0.8
• --mode: Must be set to hybrid

Oblivious Training/Unlearning Mode

The oblivious mode allows you to train or unlearn on a pre-trained forest using encrypted data from a CSV file. The same function is used for both operations; the only difference is in the label encoding:

• Training: multiply the one-hot encoded label by 1
• Unlearning: multiply the one-hot encoded label by 2

Example (Training):

cargo run --release -- --mode oblivious --forest-path ./export/iris/8/best_iris_8_4_0.95_exported.json --csv-path ./data/iris-uci/iris-train.csv --operation train --verbose

Example (Unlearning):

cargo run --release -- --mode oblivious --forest-path ./export/iris/8/best_iris_8_4_0.95_exported.json --csv-path ./data/iris-uci/iris-unlearn.csv --operation unlearn --verbose

Options:

• --mode: Must be set to oblivious
• --forest-path: Path to the pre-trained forest JSON file (required)
• --csv-path: Path to the CSV file containing data to train/unlearn (required)
• --operation: Operation type (train or unlearn). Default: train
• --output: Output directory for the updated forest. Default: ./export/
• --verbose or -v: Verbose mode

CSV Format: The CSV file should have the same format as the training data:

• Features as the first columns
• Class label as the last column
• No header row

Common Examples

Run standard mode with default settings:

cargo run --release -- --mode standard

Run hybrid mode with custom dataset and parameters:

cargo run --release -- --dataset cancer --num-trees 16 --depth 5 --mode hybrid --split 0.3 --verbose

Run standard mode with multiple forest configurations:

cargo run --release -- --dataset wine --num-trees 8,16,32 --depth 4 --mode standard --best-model-trials 5 --verbose

Run oblivious training on a pre-trained forest:

cargo run --release -- --mode oblivious --forest-path ./export/iris/8/best_iris_8_4_0.95_exported.json --csv-path ./data/iris-uci/iris-sample.csv --operation train --verbose

Run oblivious unlearning on a pre-trained forest:

cargo run --release -- --mode oblivious --forest-path ./export/iris/8/best_iris_8_4_0.95_exported.json --csv-path ./data/iris-uci/iris_to_unlearn.csv --operation unlearn --verbose

Output

The program generates:

• Exported forest files (JSON format) in the export directory
• Performance metrics and timing statistics
• Accuracy comparisons between clear and encrypted training

The output structure is:

export/
  {dataset}/
    {num_trees}/
      standard_{dataset}_{num_trees}_{depth}_{accuracy}.json  (standard mode)
      hybrid_{dataset}_{num_trees}_{depth}_{accuracy}.json  (hybrid mode)
      perf.csv  (performance metrics)
  forest_{operation}_{forest_filename}_{csv_filename}.json  (oblivious mode)