PaperlayerView exact revision on arXiv

Source-linked AI summary

ReDIL-GNN: Resynthesis Domain Incremental Learning for Circuit Graph Neural Networks

Rupesh Raj Karn, Johann Knechtel, Ozgur Sinanoglu

arXiv:2609.18595v1cs.LGcs.NE

TL;DR

Resynthesis can change circuit-graph distributions without changing labels or source identity, leaving circuit GNNs to handle shifts not captured by fixed-distribution evaluation. ReDIL-GNN evaluates domain-incremental adaptation with fixed prediction or representation heads and introduces RAI to screen new synthesis styles. Across the studied pipelines, adaptation is not uniformly beneficial, and RAI distinguishes structurally unsupported shifts from promising cases.

  • Problem

    Logic resynthesis changes circuit-graph structure while preserving labels or source identity, but fixed-distribution evaluations do not assess adaptation to new synthesis styles and retention on prior ones.

    Source: §I · p. 1

  • Method

    ReDIL-GNN evaluates sequential resynthesis-domain adaptation with fixed prediction or representation heads and uses RAI to assess adaptation need, recoverability, structural coverage, and update compatibility before training.

    Source: §I · pp. 1–2

  • Results

    Across the studied pipelines, adaptation is not uniformly beneficial, and RAI reaches 0.824 for LWF on GNN-RE original-only LUT6.

    Source: §V · p. 11

  • Takeaways & Limitations

    RAI can guide whether a new synthesis style should be learned, whether the current base model is sufficient, or whether adaptation may cause unacceptable forgetting.

    Source: §I · pp. 1–2

Abstract

from arXiv · show

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.

I. INTRODUCTION

Logic resynthesis can shift circuit-graph structure while preserving labels or source identity, so fixed-distribution evaluation does not show whether circuit GNNs can adapt to new synthesis styles and retain prior performance. ReDIL-GNN frames this as domain-incremental learning and adds RAI to screen whether a shift merits adaptation.

I. INTRODUCTION

  • Resynthesis changes gate vocabularies and graph structure while labels or source identity can remain fixed, making fixed-distribution evaluation insufficient for assessing adaptation and retention.Some shifts may already be handled by the base model, while others may be structurally unsupported or risk forgetting under naive updates.Source: §I · p. 1
  • ReDIL-GNN treats each synthesis style as a domain in a fixed semantic label space, unlike prior circuit continual-learning formulations based on synthetic task or class splits.It adapts the same prediction or representation head as new styles arrive.Source: §I · p. 1
  • RAI combines adaptation need, source-equivalence recoverability, structural coverage, and update compatibility to flag shifts as promising, unnecessary, or risky before full training.It is intended to guide whether adaptation should be invoked or avoided.Source: §I · pp. 1–2
  • The framework compares nine incremental update policies, including naive fine-tuning, retention methods, and equivalence-guided replay, across classification and representation-learning pipelines.EqReplay uses paired source-equivalent implementations across synthesis domains.Source: §I · pp. 1–2

II. RELATED WORK AND MOTIVATION FOR OURS

  • ReDIL-GNN differs from generic graph continual learning by using circuit-specific equivalence in replay and by gating costly updates with a pre-adaptation diagnostic.The framework also evaluates retention and forgetting across diverse circuit-GNN pipelines.Source: §II · p. 2

III. METHODOLOGY

ReDIL-GNN evaluates sequential resynthesis-domain shifts while retaining a model’s existing prediction head or embedding space. Its interfaces use task labels for classifiers and source-equivalence retrieval for representation models.

A. Overview of ReDIL-GNN

  • The protocol starts from a multi-domain base checkpoint, measures held-out zero-shot performance, then introduces AOXN, Lib-A, and LUT4 sequentially while retaining the same head or embedding space.The model is not given the resynthesis-domain identifier during inference.Source: A. Overview of ReDIL-GNN · p. 2
  • The setting assumes graph distributions change across resynthesis domains while labels or source identities remain fixed, so the goal is adaptation rather than adding output classes.The same prediction or representation interface is evaluated across domains.Source: A. Overview of ReDIL-GNN · p. 2

B. Model Interfaces: Classification & Representation Learning

  • Classification-based models use a shared classifier to predict task labels, with adaptation intended to learn the current graph distribution while preserving the prior decision boundary.The passage gives GNN-RE-style functional classification as an example.Source: B. Model Interfaces: Classification &… · p. 3
  • Representation-learning models use a graph-level readout and are assessed with source-equivalence retrieval measures rather than binary F1 as the primary metric.Reported measures include Recall@K, MRR, median rank, positive-pair similarity, L2 displacement, and source margin.Source: B. Model Interfaces: Classification &… · p. 3

C. Incremental Adaptation Mechanisms

ReDIL-GNN offers nine mechanisms for adapting across resynthesis domains, ranging from current-domain fine-tuning to methods that preserve prior-domain knowledge. EqReplay adds circuit-specific replay using source-equivalent implementations from adjacent domains.

D. Stage-Wise Training and Evaluation

Each incoming resynthesis domain is processed in stages: the selected mechanism updates the model, then evaluation covers the current and previously observed domains. The protocol separates current-domain adaptation from retention and uses task-native metrics for classification and representation-learning models.

D. Stage-Wise Training and Evaluation

  • After each domain update, evaluation across the current and previously observed domains forms a stage-by-domain matrix for final, worst-domain, and forgetting measurements.The update may use allowed memory, teacher, importance estimates, or source-equivalent replay examples.Source: D. Stage-Wise Training and Evaluation · pp. 3–4
  • Forgetting is the drop from a domain's best previously achieved score to its score after the final incremental stage.Current-domain performance measures plasticity, while prior-domain performance measures stability.Source: D. Stage-Wise Training and Evaluation · p. 4
  • Classification models use metrics such as macro-F1, balanced accuracy, or task-specific detection performance, while representation models primarily use MRR, Recall@K, or source margin.This lets the protocol compare both model families without forcing them into one metric type.Source: D. Stage-Wise Training and Evaluation · p. 4, Table III · p. 4

E. Resynthesis Adaptability Index

The Resynthesis Adaptability Index (RAI) uses four normalized signals to estimate whether adapting to an incoming synthesis domain is needed, recoverable, structurally covered, and compatible with retention. Its conservative score guides whether to skip adaptation, pursue it, or use retention-heavy safeguards.

E. Resynthesis Adaptability Index

  • RAI combines adaptation need, source-equivalence recoverability, structural coverage, and update compatibility as a geometric mean of normalized terms.The geometric form is conservative: a low term can drive the score toward zero, while values near one indicate a strong candidate for safe and useful adaptation.Source: E. Resynthesis Adaptability Index · pp. 4–5
  • Adaptation need is high when the frozen base model performs poorly on the incoming-domain probe, while a low score indicates that it already generalizes to that style.The normalized zero-shot score can be macro-F1, balanced accuracy, task-specific F1, Recall@1, or MRR depending on model type.Source: E. Resynthesis Adaptability Index · p. 4
  • Source-equivalence recoverability measures whether the base model still recognizes different implementations of the same source circuit as related.Embedding models use the similarity gap between the correct equivalent pair and the closest incorrect source; classification models use prediction consistency.Source: E. Resynthesis Adaptability Index · p. 4
  • Structural coverage compares descriptors of the incoming domain with those of base domains; high coverage indicates closeness to a previously seen style.Descriptors may include graph statistics such as node and edge counts, gate histograms, fan-in/fan-out, depth, or frozen-model embedding summaries.Source: E. Resynthesis Adaptability Index · p. 4
  • Update compatibility estimates whether an incoming-domain probe update agrees with the base-domain retention objective; low compatibility indicates possible forgetting.The probe gradient is compared with the gradient on the base-domain retention probe.Source: E. Resynthesis Adaptability Index · pp. 4–5
  • When need is low, adaptation can be skipped; when need is high but RAI is low, retention-heavy mechanisms are preferred over naive fine-tuning.The structural context includes statistics for eight GNN-RE resynthesis outcomes and mean-over-eight summaries across reproduced SOTA works.Source: E. Resynthesis Adaptability Index · p. 5, Table IV · p. 5, Table V · p. 5

A. Resynthesis Statistics

Resynthesis changes circuit graph structure beyond superficial gate renaming, and classification-based Domain-IL results show that adaptation must balance incoming-domain performance with retention. Multi-domain training stabilizes the starting point but does not remove mechanism-dependent trade-offs.

A. Resynthesis Statistics

  • Resynthesis produces substantial structural changes, including large edge reductions for LUT mappings and component and gate-composition changes across classification and embedding pipelines.Reported graph statistics include topology, depth, fan-in/fan-out, density, components, gate-type entropy, and drift measures.Source: A. Resynthesis Statistics · p. 5

B. Domain-IL Results for Classification-Based SOTA GNNs

  • The Original-only GNN-RE experiment is a lower-bound deployment scenario: performance varies across resynthesis styles and adaptation mechanisms.The base model is adapted sequentially across ABC-rewrite, AON, AOXN, NAND, Lib-A, Lib-D, LUT4, and LUT6.Source: B. Domain-IL Results for… · p. 6
  • The main GNN-RE experiment starts from a multi-domain base and learns the held-out AOXN → Lib-A → LUT4 stream, tracking task metrics and base-domain forgetting.Compared with the Original-only baseline, the multi-domain base has a more stable starting point.Source: B. Domain-IL Results for… · p. 6
  • Replay, distillation, and equivalence-aware mechanisms generally offer a better stability–plasticity trade-off than naive adaptation, but no mechanism dominates all classification models.Current-domain performance and retention do not improve uniformly across mechanisms.Source: B. Domain-IL Results for… · p. 6

C. Domain-IL Results for Embedding-Based SOTA GNNs

Embedding-based Domain-IL results vary across models: FGNN2 remains stable on the evaluated stream, while other encoders show differing retrieval robustness and forgetting. The results support evaluating source-equivalence retrieval with multiple metrics rather than a single score.

C. Domain-IL Results for Embedding-Based SOTA GNNs

  • FGNN2 retrieval metrics remain close to saturation for most mechanisms through AOXN, Lib-A, and LUT4, with limited base-domain retrieval loss after LUT4.The figure tracks MRR, Recall@1, Recall@5, Recall@10, mechanism distributions, and base-retention MRR forgetting.Source: C. Domain-IL Results for… · pp. 6–7
  • After LUT4, embedding-based models differ in robustness: FGNN2 remains strongest, Netlist/Circuit GNN is moderate, and DeepGate-family models are more sensitive to the stream.PolarGate's high positive-pair similarity for some mechanisms does not ensure strong Recall@1 or margin when hard negatives remain close.Source: C. Domain-IL Results for… · p. 7
  • Replay and hybrid replay-distillation mechanisms generally reduce retrieval degradation, but EqReplay does not uniformly dominate every metric.Retention remains mechanism-dependent across the consolidated embedding results.Source: C. Domain-IL Results for… · p. 7
  • Resynthesis retrieval should be assessed with Recall@K, MRR, margin, distance, similarity, and forgetting rather than one score alone.These measures capture source-identification success, hard-negative separation, embedding alignment, and retention.Source: C. Domain-IL Results for… · p. 7

D. RAI Results for Classification-based SOTA GNNs

Classification-side RAI varies with incoming-domain need, recoverability, coverage, and mechanism compatibility. In the Original-only GNN-RE setting, the index ranges from an unsupported shift to promising adaptation cases.

D. RAI Results for Classification-based SOTA GNNs

  • In the multi-domain classification setting, GNN-RE's RAI is 0.558 on AOXN, 0.668 for EQREPLAY on Lib-A, and 0.610 for DER++ on LUT4.For GNN4IP at LUT4, EWC scores 0.582 versus ER at 0.435; AppGNN's LUT4 scores are close, from 0.793 to 0.802.Source: D. RAI Results for… · p. 8
  • RAI is 0.001 for every mechanism on Original-only GNN-RE ABC-rewrite because structural coverage is zero despite high adaptation need (N = 0.798).In later stages, EQREPLAY reaches 0.743 on AOXN, while LWF reaches 0.796 on LUT4 and 0.824 on LUT6.Source: D. RAI Results for… · p. 9
  • RAI is adaptation guidance, not a performance leaderboard: tied scores reflect domain-level factors, while separated scores can reflect mechanism-specific compatibility and retention risk.The index can guide whether to skip an update or use replay, regularization, distillation, or EqReplay protection.Source: D. RAI Results for… · p. 9

E. RAI Results for Embedding-based SOTA GNNs

Embedding-side RAI reflects structural support, source-equivalence recoverability, and mechanism compatibility, while the paper also evaluates whether RAI tracks adaptation gain and forgetting. The reported results position RAI as a guide to update or defer decisions, not a universal predictor.

E. RAI Results for Embedding-based SOTA GNNs

  • DeepGate-family LUT4 RAI scores collapse toward zero when structural coverage is zero, despite high adaptation need; FGNN2's tied AOXN RAI is 0.328 with low need (N = 0.020).Other selected cases include MAS at 0.717 for PolarGate LUT4, LWF at 0.581 for NetlistGNN LUT4, and EQREPLAY for FGNN2's later stages.Source: E. RAI Results for Embedding-based… · p. 9
  • When coverage and source-equivalence recoverability are strong, EqReplay, MAS, or LwF can be suitable updates depending on the model; high need alone is insufficient when coverage collapses.The paper recommends treating low-coverage LUT cases for DeepGate-family models as unsupported and seeking more source-equivalent views or stronger retention constraints.Source: E. RAI Results for Embedding-based… · pp. 9–10

F. RAI Predictiveness and Decision Rules

  • RAI decision thresholds are empirical operating rules, not universal constants, for skipping updates, requesting more source-equivalent views, or selecting a mechanism.Low need suggests existing coverage, low structural coverage flags unsupported shifts, and low compatibility warns of retention conflict.Source: F. RAI Predictiveness and Decision Rules · p. 10
  • RAI has positive adaptation-gain correlations of ρ = 0.41 and 0.45 across classification SOTA and embedding SOTA; compatibility is the clearest retention signal, with ρ(C, −F) = 0.62 across rows with forgetting available.RAI is not a universal scalar predictor; the tie-aware Top-1 rate is 56.2% and median regret is 0.0.Source: F. RAI Predictiveness and Decision Rules · p. 10
  • RAI's actionable output is a choice among reusing the current model, applying protected adaptation, or deferring until the new synthesis style is better supported.Its full score identifies adaptation benefit and unsupported cases, while compatibility is most directly tied to forgetting risk.Source: F. RAI Predictiveness and Decision Rules · p. 10

G. RAI Computational Cost

RAI provides a low-overhead pre-adaptation decision for ReDIL-GNN, while the conclusion shows that adaptation support varies sharply across resynthesis shifts.

G. RAI Computational Cost

  • RAI completes within minutes, while full adaptation or retraining requires hours, making its pre-adaptation decision overhead practical before committing to expensive updates.The measured comparison includes probe construction and RAI scoring versus incremental optimization and evaluation.Source: G. RAI Computational Cost · p. 10, Table XI · p. 11

V. CONCLUSION

  • ReDIL-GNN evaluates circuit GNNs under sequential, functionality-preserving resynthesis shifts across classification-based and embedding-based models.The conclusion describes the framework as covering nine adaptation mechanisms.Source: §V · p. 10
  • RAI ranged from 0.001 for GNN-RE original-only ABC-rewrite to 0.824 for LWF on GNN-RE original-only LUT6, warning when adaptation is structurally unsupported.The results indicate that adaptation mechanisms become useful only when adaptation need, source recoverability, structural coverage, and update compatibility are jointly favorable.Source: §V · p. 11
  • Future work would extend RAI into an active controller that can skip adaptation, select retention-heavy updates, allocate replay memory, or request more source-equivalent views.The proposed controller would make these choices before updating the circuit GNN.Source: §V · p. 11

Related papers

Loading 2609.18595v1…