VERIFY: Key Assumption Validation and Spectral Graph Learning for Oracle-Less Logic Unlocking

This is the implementation for the VERIFY research project.

Project Structure

.
├── data/                  # Original circuit datasets
│   ├── TRLL/              # TRLL locked circuit data
│   ├── TRLL+/             # TRLL+ locked circuit data
│   └── TroMUX/            # TroMUX locked circuit data
├── data_part/             # Datasets with standard features only
├── rewire/                # Circuit rewiring module based on Ollivier-Ricci curvature
│   ├── netlists/          # Original netlist circuit data
│   ├── GraphRicciCurvature/  # Ricci curvature computation library
│   ├── preprocessing/     # Data preprocessing scripts
│   └── run_netlists_rewire.py  # Main rewiring script
├── main.py                # Main program entry
├── BWGNN.py               # Beta Wavelet Graph Neural Network implementation
├── bonGNN.py              # Bernstein polynomial GNN implementation
├── cheGNN.py              # Chebyshev polynomial GNN implementation
└── dataset.py             # Dataset loading and processing module

Data Format

Each circuit dataset (e.g., data/TRLL/c1355/0/) contains the following files:

• edges.txt: Graph edge connections
• feat.txt: Node feature matrix
• keys.txt: Indices of key nodes
• label.txt: Node labels (0 for normal nodes, 1 for key nodes)

File Descriptions

main.py

Main program entry responsible for loading data, constructing graphs, training models, and evaluating results. Includes command-line argument parsing, data preprocessing, model training, and evaluation.

BWGNN.py

Implements Beta Wavelet Graph Neural Network, using Beta wavelet basis functions as graph filters, effectively capturing both local and global structural information in graphs.

bonGNN.py

Implements Bernstein polynomial-based Graph Neural Network, using Bernstein polynomials as graph filters, suitable for processing non-stationary graph signals.

cheGNN.py

Implements Chebyshev polynomial-based Graph Neural Network, approximating spectral graph convolutions with Chebyshev polynomials for more efficient graph signal processing.

dataset.py

Dataset loading and processing module, supporting various dataset formats.

rewire/

Circuit rewiring module based on Ollivier-Ricci curvature. This module performs graph rewiring on circuit netlists to enhance robustness against logic locking attacks. Key features include:

• Ollivier-Ricci Curvature: Computes curvature metrics for graph edges
• Circuit Support: Supports various benchmark circuits (b11, b14, b17, c1355, c1908, c2670, c3540, c6288, etc.)
• Data Processing: Converts between different graph data formats and applies rewiring transformations

Dependencies

Core Model

• pytorch 1.9.0
• dgl 0.8.1
• sympy
• argparse
• sklearn

Rewire Module

• pytorch
• torch-geometric
• networkx
• scipy
• numpy

How to Run

Main Model (VERIFY)

python main.py --dataset amazon --train_ratio 0.4 --hid_dim 64 \
--order 2 --homo 1 --epoch 100 --run 1 --lockedName TRLL --circuit c1355 \
--test_start 449 --val_start 400 --lenss 450 --dataSetType data

Parameter Descriptions

• --dataset: Dataset name, e.g., amazon, yelp
• --train_ratio: Training set ratio, default 0.4
• --hid_dim: Hidden layer dimension, default 64
• --order: Order C parameter in Beta Wavelet, default 2
• --homo: 1 for BWGNN(Homo), 0 for BWGNN(Hetero)
• --epoch: Maximum number of training epochs, default 100
• --run: Number of runs, default 1
• --lockedName: Name of the lock, default "TRLL"
• --circuit: Circuit name, default "c1355"
• --test_start: Starting index for test set, required
• --val_start: Starting index for validation set, required
• --lenss: Total data length, required
• --dataSetType: Dataset type directory, required

Circuit Rewiring (Rewire Module)

cd rewire
python run_netlists_rewire.py \
    --circuit_type c1355 \
    --num_circuits 10 \
    --num_iterations 3 \
    --borf_batch_add 20 \
    --borf_batch_remove 10

Rewire Parameter Descriptions

• --circuit_type: Circuit type name (e.g., b11, b14, c1355, c1908, c2670, c3540, c6288)
• --num_circuits: Number of circuit instances to process (starting from 0)
• --num_iterations: Number of rewiring iterations
• --borf_batch_add: Number of edges to add per iteration
• --borf_batch_remove: Number of edges to remove per iteration
• --netlists_dir: Input netlists directory (default: netlists)
• --output_dir: Output directory (default: graphData/rewired)

Acknowledgement

The band-pass filter implementation in VERIFY utilizes the beta-wavelet codes from the following paper: J. Tang, J. Li, Z. Gao, and J. Li, Rethinking graph neural networks for anomaly detection, ICML 2022. Thanks to the authors for making their code available.