How sudiptomaity2005/trojannet-gnn works
TrojanNet-GNN detects and localizes hardware Trojans in gate-level netlists without requiring golden references. Its main workflow parses circuits into graphs, simulates behavioral activity, forms graph features, and applies GraphSAGE embeddings with cost-sensitive classifiers; evaluation uses cross-family leave-one-circuit-out testing. The repository also includes TARMAC trigger-vector generation as a separate capability. Some model and evaluation implementations were not excerpted, so those branches reflect README descriptions rather than inspected call paths.
Main entry points: parser.py Netlist graph parser · rs232_pipeline.py Circuit dataset pipelines · colab_feature_engineering.py Behavioral and topology features · colab_gnn_ml_classifiers.py Hybrid feature vector
Components 17
Circuit preparation
- Netlist graph parser
src/data_preparation/parser.py - Logic simulator
src/data_preparation/simulator.py - Circuit dataset pipelines
src/data_preparation/rs232_pipeline.py - Benchmark graph datasets
src/data_preparation/generate_gnn_datasets.py
Trojan detection
- Behavioral and topology features
src/models/colab_feature_engineering.py - GraphSAGE encoder
src/models/colab_gnn_script.py - Hybrid feature vector
src/models/colab_gnn_ml_classifiers.py - Cost-sensitive classifiers
src/models/colab_gnn_ml_classifiers.py - Trojan gate localization
src/models/localization_pipeline.py
Evaluation
- Cross-family LOCO evaluation
src/models/run_strict_loco_6classifiers.py - Detection metrics
scripts/colab_multiseed_loco_eval.py
TARMAC triggers
- TARMAC trigger generator
TARMAC_trigger_activation/tarmac.py - Optimized TARMAC
TARMAC_trigger_activation/tarmac_optimized.py - TARMAC experiments
TARMAC_trigger_activation/run_experiments.py - Trigger test vectors
Other components
- Security researcher
- Gate-level netlists
Connections 19
- Security researcher → Gate-level netlists: provides circuits
- Gate-level netlists → Netlist graph parser: parsed into graph
- Netlist graph parser → Logic simulator: supplies graph
src/data_preparation/rs232_pipeline.py - Circuit dataset pipelines → Netlist graph parser: builds circuit graph
src/data_preparation/rs232_pipeline.py - Circuit dataset pipelines → Logic simulator: simulates circuits
src/data_preparation/rs232_pipeline.py - Circuit dataset pipelines → Benchmark graph datasets: exports features
src/data_preparation/rs232_pipeline.py - Benchmark graph datasets → Behavioral and topology features: provides graph data
- Behavioral and topology features → GraphSAGE encoder: encodes topology
- Behavioral and topology features → Hybrid feature vector: combines features
- GraphSAGE encoder → Hybrid feature vector: provides embeddings
- Hybrid feature vector → Cost-sensitive classifiers: classifies gates
src/models/colab_gnn_ml_classifiers.py - Cost-sensitive classifiers → Trojan gate localization: predicts Trojan gates
- Trojan gate localization → Cross-family LOCO evaluation: evaluated across folds
- Cross-family LOCO evaluation → Detection metrics: reports metrics
- Security researcher → TARMAC experiments: runs experiments
TARMAC_trigger_activation/run_experiments.py - TARMAC experiments → Optimized TARMAC: calls generator
TARMAC_trigger_activation/run_experiments.py - TARMAC trigger generator → Netlist graph parser: uses netlist parser
TARMAC_trigger_activation/tarmac.py - TARMAC trigger generator → Trigger test vectors: generates vectors
- Optimized TARMAC → Trigger test vectors: generates vectors