A research playground for verifiable dropout: generating and checking zero-knowledge proofs that a dropout layer was sampled and applied correctly during neural network training. The project combines Python reference code with a minimal RISC Zero zkVM workspace that proves deterministic dropout over quantized tensors.

• verifiable_dropout.py: Core Python module implementing VRF-backed dropout masks, Merkle-based auditing, zkVM integration helpers, and a demo runner.
• run_experiments_final.py: Batch experiment harness that compares local hashes against zkVM outputs across dropout probabilities and shapes.
• debug_2.py: Diagnostic script for probing rounding/quantization rules against zkVM receipts.
• make_plots.py: Utilities for turning experiment CSV results into publication-ready figures.
• vdo_zkvm/: Minimal Rust workspace containing zkVM guest (methods/guest), embedding crate (methods), and host CLI for proving (dropout_cli) and verifying (dropout_verify_cli).

Install the Python dependencies into a virtual environment (Python 3.10+ recommended):

python -m venv .venv
source .venv/bin/activate
pip install torch cryptography numpy pandas matplotlib tqdm

verifiable_dropout.py and the experiment scripts expect PyTorch and the cryptography package for Ed25519-based VRF signatures.

The zkVM side is optional for modes A/B but required for mode C proofs.

cd vdo_zkvm
# Fast dev mode (no real proving)
RISC0_DEV_MODE=1 cargo test -p host --release
# Real proving
cargo test -p host --release

Successful builds produce:

• vdo_zkvm/target/release/dropout_cli – proves dropout and emits a receipt/journal hash.
• vdo_zkvm/target/release/dropout_verify_cli – verifies receipts and replays the journal.

verifiable_dropout.py contains a self-contained demo that:

1. Builds a simple MLP with a VerifiableDropout layer.
2. Generates a VRF seed for the layer.
3. Calls the zkVM host (dropout_cli) to produce a receipt.
4. Verifies the receipt with dropout_verify_cli.

# From the repository root (after building the zkVM binaries)
export ZKVM_DROPOUT_CLI="$(pwd)/vdo_zkvm/target/release/dropout_cli"
export ZKVM_DROPOUT_VERIFY_CLI="$(pwd)/vdo_zkvm/target/release/dropout_verify_cli"
python verifiable_dropout.py

mode controls the proof style:

• A: VRF packet only (fast, no audit).
• B: Merkle roots of pre/post tensors with block sampling proofs.
• C: Full zkVM receipt verifying quantized dropout output and mask.

• run_experiments_final.py sweeps dropout probabilities and tensor shapes, invoking the zkVM binaries and writing CSV summaries. Key options include --p-values, --shapes, and --risc0-dev-mode to skip heavy proving during development.
• debug_2.py offers lower-level reproduction of zkVM hashes with different rounding rules to match guest behavior.
• make_plots.py consumes CSV outputs (e.g., from run_experiments_final.py) and writes PNG/PDF figures to an output directory.

Example (dev-mode, small sweep):

# Build binaries first (see above)
python run_experiments_final.py --p-values 0.1 0.2 0.5 --shapes 8x16 --risc0-dev-mode --emit-receipt
python make_plots.py --exp1-csv results/exp1.csv --outdir plots

• VerifiableDropout uses deterministic seeds derived from a VRF over (model_id, step, batch_id, nonce, layer_id), ensuring reproducible masks for both prover and verifier.
• Quantization/rounding logic in verifiable_dropout.py mirrors the Rust guest: int32 tensors scaled by p_den/(p_den-p_num) with half-away-from-zero rounding before clamping.
• Set RISC0_DEV_MODE=1 in the environment for faster iterations when calling dropout_cli from Python wrappers.
• If the zkVM binaries are in a non-default location, pass zkvm_bin/verify_bin arguments to the Python helpers or set the environment variables used in the demo.