Resynthesis domain learning improves circuit graph neural network adaptation
ReDIL-GNN: Resynthesis Domain Incremental Learning for Circuit Graph Neural Networks
Machine LearningNeural and Evolutionary Computing
Summary
Changes in how computer circuits are built can confuse models that analyze them since the circuit’s structure changes but its function stays the same. The authors created a method called ReDIL-GNN that helps models adapt to these changes by deciding when and how to update themselves. They also designed a score, the Resynthesis Adaptability Index, to predict if adapting to a new circuit style will help or hurt the model. Their approach helps keep models accurate over time by smartly choosing when to learn from new circuit designs.
What this means in practice
- •For hardware security teams: Maintain reliable circuit analysis models that stay accurate despite changes in circuit design methods over time.
- •For chip design engineers: Use metrics to predict if updating a model for new circuit synthesis styles will improve analysis without losing earlier knowledge.
Authors
Rupesh Raj Karn, Johann Knechtel, Ozgur Sinanoglu
Abstract
Logic resynthesis preserves circuit functionality while changing gate vocabulary, topology, and structural statistics, creating domain shift for circuit graph neural networks (GNNs) without changing task labels. To study this setting, we introduce ReDIL-GNN, a resynthesis domain-incremental learning framework that adapts a fixed prediction or representation head as new synthesis styles arrive and evaluates retention on all previously observed domains. Because not every shift should be adapted blindly, ReDIL-GNN further introduces the Resynthesis Adaptability Index (RAI), a pre-adaptation score that combines adaptation need, source-equivalence recoverability, structural coverage, and update compatibility. We evaluate supervised hardware-security tasks and representation-learning models using task-native metrics for classifiers and source-equivalence retrieval metrics for embedding models, comparing naive fine-tuning with LwF, Online EWC, MAS, ER, A-GEM, DER++, ER+LwF, and equivalence-guided replay. Across the studied pipelines, RAI separates unsupported shifts from promising updates, ranging from 0.001 for a structurally uncovered GNN-RE ABC-rewrite shift to 0.824 for the best original-only GNN-RE adaptation case. In practice, ReDIL-GNN turns resynthesis-aware circuit learning into a deployment control loop: RAI screens each new synthesis flow before update, guiding whether to reuse the current model, apply retention-aware adaptation, or defer adaptation until the shift is better supported.