cynthi8 / gnnunlock

Reproduction of https://github.com/DfX-NYUAD/GNNUnlock

GNNUnlock

Reproduction of https://github.com/DfX-NYUAD/GNNUnlock

This project reproduces the steps of Lilas Alrahis et al. in GNNUnlock up to evaluating the GNN (Graph Neural Network).

SFLL-HD explains that I was able reproduce the success of the original authors for netlists locked with SFLL-HD and Cross Compilation explains an essential shortcoming with GNNUnlock.

SFLL-HD

I chose to reproduce the setup in ANTI_SAT_DATASET_c7552 except with SFLL-HD_2. This setup uses 24 netlists of c2670 and c5315 as training graphs, 9 netlists of c3540 as validation graphs, and 12 netlists of c7552 as testing graphs (total: 45 netlists). Each netlist is locked with various key sizes and different values and are contained in Datasets/SFLL_HD_2.

The training log for this setup is shown in Datasets/SFLL_HD_2/log_training.txt. The GNN performance on the test graphs is shown below:

Confusion Matrix

True \ Predicted	restore	perturb	main
restore	888	0	0
perturb	0	654	22
main	15	5	8518

F1 Score by Class

main	perturb	restore
0.992	0.980	0.998

Cross Compilation

In GNNUnlock, training, validation, and testing are all done on non-cross-compiled netlists. Any realistic implementation of logic locking would flatten the design heirarchy before compiling which would further obfuscate the main design/pertub/restore node classes.

All netlists were compiled with Synthesizer/dc_shell_script.tcl, while Synthesizer/dc_shell_script_cross_compile.tcl would be a more "realistic" implementation.

Setup Steps:

1. Run source setup_conda_env.sh to setup the anaconda enviroment
2. Run setup_GraphSAINT.sh to clone GraphSAINT and build it

Remain in the gnnunlock-env anaconda enviroment while working within this repository.

Reproduction Steps:

The bolded directories at each step contain the relevant files.

1. Obtain Benchmarks in Bench format. These are already saved in Benchmarks
2. Convert to Verilog. python convert_bench_files_to_verilog.py
3. Lock with SFLL and AntiSAT. python lock_netlists.py
4. Synthesize with Design Compiler™. Examples are already saved in SynthesizedLockedNetlists

cd Synthesizer/ 
python synthesize_netlists.py

5. Assign roles (train, validation, test). Examples are saved in Datasets
6. Convert to graph and extract features.

python convert_netlists_to_graph.py

7. Train GNN.

./trainGNN.sh

Epochs Trained

Adjust the phase: end parameter in DATE21.yml to increase the training time. Data were collected with end: 2000 to be consistant with GNNUnlock.

gnnunlock's Issues

Wrong Feature Extraction

Dear Erin,

Please check our gate-level Netlist parser in https://github.com/DfX-NYUAD/GNNUnlock to get our results.
Your feature extraction is not consistent with our proposed extraction in DATE'21 paper.
Lilas