MicroSim: Contrastive Graph Neural Networks for Binary Code Similarity Detection
Authors: Moussa Moussaoui, Tarik Houichime, Abdelghani Souhar, Abdelalim Sadiq
Organizations: Computer Science Research Laboratory(LaRI), Faculty of Science, Ibn Tofail University, Kenitra, 14000, Morocco · Meridian Team, LyRICA Laboratory, School of Information Sciences, Rabat 10100, Morocco
We present MicroSim, a system for binary code similarity detection that represents each function as a graph built from IDA Pro Hex-Rays microcode and compares such graphs with a contrastive Siamese graph neural network (GNN). We evaluate on the TREX cross-optimization benchmark, a corpus of 849,072 function-level microcode graphs spanning optimization levels -O0 through -O3, under a strict class-disjoint split in which no function key is shared between train and test. Under this protocol, MicroSim reaches 0.9972 area under the receiver operating characteristic curve (ROC-AUC) and 0.9762 Matthews correlation coefficient (MCC), outperforming Gemini (0.9332 ROC-AUC), SAFE (0.9752), and jTrans (0.9740) retrained on the same split, as well as a structure-free Bag-of-Words (BoW) baseline (0.9484). We analyze the gap between pair classification and ranked retrieval in graph transformers, showing that Cross-domain Similarity Local Scaling (CSLS) under unified tie-resolved evaluation raises GraphGPS mean reciprocal rank (MRR) from 0.2542 to 0.2559 (R@1 = 0.1639, R@10 = 0.4561), strictly satisfying the theoretical MRR upper bound. Finally, over 5 seeds with equal compute, the pairwise cosine-embedding loss attains a slightly higher ROC-AUC (Δ = -0.0256, 95% confidence interval [-0.3282, +0.2770]), whereas the supervised-contrastive information noise-contrastive estimation (InfoNCE) loss is significantly better on thresholded classification and ranked retrieval.
Figures & tables
Fig. 1 : MicroSim: a control-/data-flow microcode graph distilled to a 128-dimensional embedding for binary similarity, with the (A) GraphGPS encoder upgrade and (B) GraphCL pre-training extensions.
Parameter
GCN
GraphGPS (Transformer)
Layers
2
3
Hidden size ⋅ heads
128 ⋅ —
128 ⋅ 4
Positional encoding
—
RWSE, length 8
Pooling ⋅ embedding size
mean ⋅ 128
mean ⋅ 128
Block features ⋅ block cap
64-d hashed ⋅ none
64-d hashed ⋅ 256
Batch ( B anchors, K neg/anchor)
B=64 , K=8
B=64 , K=8
TABLE I : Architecture and Implementation Details of MicroSim Encoders
Model
ROC-AUC
F1
MCC
MRR
R@1
R@10
Bag-of-Words (BoW)
0.9484
0.8744
0.7775
0.2114
0.1375
0.3657
Gemini [ 3 ]
0.9332
0.8486
0.7318
0.1553
0.0961
0.2749
SAFE [ 4 ]
0.9752
0.9237
0.8659
0.2134
0.1377
0.3716
jTrans [ 5 ]
0.9740
0.9262
0.8687
0.2124
0.1366
0.3711
MicroSim (GCN)
0.9905
0.9603
0.9288
0.2177
0.1380
0.3860
MicroSim (GraphGPS)
0.9972
0.9869
0.9762
0.2542
0.1618
0.4568
TABLE II : MicroSim and Baselines on the TREX Cross-Optimization Benchmark (Test Split, N=126,769 Functions)
Fig. 2 : Empirical CDF of positive match rank rq on the TREX test split ( N=126,769 functions) under unified tie-resolved evaluation. Post-hoc CSLS and score fusion shift the distribution toward top ranks.
Method / Variant
Type
MRR
R@1
R@10
GraphGPS (raw, τ=0.1 )
baseline
0.2542
0.1618
0.4568
+ Mean-centering & whitening
post-hoc
0.2543
0.1623
0.4565
+ CSLS re-ranking ( k=10 )
post-hoc
0.2559
0.1639
0.4561
+ Score fusion (GCN + GraphGPS)
post-hoc
0.2499
0.1594
0.4480
+ Two-stage (top-100 → GCN)
post-hoc
0.2348
0.1468
0.4241
GraphGPS ( τ=0.05 )
retrain
0.9457±0.0615
0.9321±0.0575
0.9774±0.0575
TABLE III : Retrieval Diagnostics and Re-Ranking Improvements on the TREX Test Split
Setting
Model
ROC-AUC
MRR
R@10
GCC ↔ Clang
Bag-of-Words
0.8912±0.0014
0.1842±0.0050
0.2895±0.0080
MicroSim (GCN)
0.9684±0.0011
0.2265±0.0050
0.3482±0.0099
MicroSim (GraphGPS)
0.9846±0.0007
0.2648±0.0058
0.4115±0.0075
Plain ↔ O-LLVM
Bag-of-Words
0.8426±0.0019
0.1418±0.0041
0.2264±0.0061
MicroSim (GCN)
0.9345±0.0014
0.1946±0.0047
0.3018±0.0073
MicroSim (GraphGPS)
0.9672±0.0009
0.2412±0.0061
0.3765±0.0097
TABLE IV : Zero-Shot Generalization to Clang and Obfuscated Builds
Metric
Pairwise
InfoNCE
Δ [95% CI]
p (Holm)
ROC-AUC
0.9929±0.0005
0.9834±0.0007
−0.0095[−0.0102,−0.0088]
<0.001
F1
0.9428±0.0014
0.9640±0.0011
+0.0212[+0.0198,+0.0227]
<0.001
MCC
0.8948±0.0026
0.9358±0.0020
+0.0410[+0.0383,+0.0438]
<0.001
MRR
0.2284±0.0020
0.2642±0.0018
+0.0358[+0.0336,+0.0380]
<0.001
R@1
0.1213±0.0015
0.1431±0.0014
+0.0218[+0.0201,+0.0235]
<0.001
R@5
0.2513±0.0022
0.2979±0.0021
+0.0466[+0.0441,+0.0491]
<0.001
TABLE V : Compute-Matched Objective Ablation on the TREX Test Split ( N=126,769 Graphs, Mean ± 95% CI over 5 Seeds; Δ = InfoNCE minus Pairwise)
Encoder
Variant
ROC-AUC
MCC
MRR
R@1
R@10
GCN
Full (CFG ∪ DFG)
0.9905±0.0005
0.9288±0.0035
0.2486±0.0059
0.1406±0.0044
0.3785±0.0074
Without DFG (CFG only)
0.9841±0.0008
0.8964±0.0045
0.2294±0.0061
0.1284±0.0037
0.3518±0.0070
Without CFG (DFG only)
0.9786±0.0009
0.8712±0.0057
0.2168±0.0064
0.1196±0.0036
0.3342±0.0069
Without block features
0.5413±0.0029
0.0842±0.0051
0.0215±0.0021
0.0084±0.0013
0.0364±0.0029
Features only (no edges)
0.9614±0.0010
0.8094±0.0074
0.2398±0.0059
0.1412±0.0032
0.3629±0.0082
GraphGPS
Full (CFG ∪ DFG)
0.9973±0.0004
0.9756±0.0023
0.2894±0.0070
0.1656±0.0045
0.4466±0.0097
TABLE VI : Ablation of Control Flow, Data Flow, and Block Features (Mean ± 95% CI over 3 Seeds)
Fig. 3 : Test MRR and empirical token hash collision rate across block feature hash dimension d∈{16,32,64,128,256} . d=64 achieves the best balance between token collision rate and parameter efficiency.
Size Bin (Blocks)
Test Share
GCN MRR
GraphGPS MRR
GCN R@10
GraphGPS R@10
1 – 16
73.84%
0.2584±0.0014
0.2942±0.0014
0.3912±0.0020
0.4538±0.0020
17 – 64
22.41%
0.2268±0.0024
0.2815±0.0024
0.3495±0.0033
0.4362±0.0033
65 – 256
3.54%
0.1892±0.0062
0.2468±0.0062
0.3042±0.0084
0.3945±0.0084
>256
0.21%
0.1425±0.0109
0.1984±0.0109
0.2368±0.0149
0.3270±0.0149
TABLE VII : Retrieval Performance by Function Size Bin on the TREX Test Split
Graph contrastive learning trains graph representations by aligning related nodes or graph views while separating structurally or semantically different examples.
Binary function embedding models are trained to encode the semantics of binary code in such a way that they can be generalized to a variety of reverse engineering tasks, such as binary code search, vulnerability detection, or malware classification. While many models only take the function in question as contextual input, there have been successful attempts to improve function embeddings by leveraging information from the call graph. In this study, we dissect the implications of these embedding refinements. We conduct experiments using a range of graph-based models on the embeddings generated by two state-of-the-art binary function embedding models. Integrating inter-procedural context, we show that improvements on binary code similarity detection (BCSD) will not necessarily generalize to downstream tasks, neither of semantic nor of syntactic nature. More generally, we find that optimizing for semantic similarity tasks correlates with worse performance on syntactic tasks. By conducting an explanatory analysis on the dataset, we find that the call graph-based enhancements significantly enhance the robustness of embeddings, particularly in scenarios where the initial models struggle. Furthermore, we observe that the added context is more beneficial for namespace-related functions than for those focused on individual logic, confirming that the call graph can be leveraged most effectively in context-dependent scenarios.
Samuel Valenzuela, Johannes Kinder
Ludwig-Maximilians-Universität München · Munich Center for Machine Learning · Center for Digital Technology and Management +1
Binary code representation learning is a fundamental problem in software security and reverse engineering. Existing methods mainly learn function-level embeddings that capture coarse-grained semantic relationships between binary functions, but they largely ignore fine-grained instruction-level correspondences. This limitation misses valuable supervision signals available from compiler debug information, which can support the learning of more accurate and interpretable binary code representations. We propose to leverage instruction alignment knowledge to further improve binary code representation learning. Our preliminary study reveals that models finetuned for function-level binary code similarity exhibit substantially better instruction alignment than their pre-trained model, suggesting a strong correlation between instruction alignment and function-level embedding quality. Motivated by this observation, we design a training approach that explicitly incorporates instruction alignment as an auxiliary training objective. Our experiments show that instruction alignment training improves retrieval accuracy and provides more discriminative signal for the model's similarity judgments.
Huaijin Wang, Shuai Wang
Shandong University · Hong Kong University of Science and Technology
Binary Function Similarity Detection (BFSD) is a core problem in software security, supporting tasks such as vulnerability analysis, malware classification, and patch provenance. In the past few decades, numerous models and tools have been developed for this application; however, due to the lack of a comprehensive universal benchmark in this field, researchers have struggled to compare different models effectively. Existing datasets are limited in scope, often focusing on a narrow set of transformations or types of binaries, and fail to reflect the full diversity of real-world applications. We introduce EXHIB, a benchmark comprising five realistic datasets collected from the wild, each highlighting a distinct aspect of the BFSD problem space. We evaluate 9 representative models spanning multiple BFSD paradigms on EXHIB and observe performance degradations of up to 30% on firmware and semantic datasets compared to standard settings, revealing substantial generalization gaps. Our results show that robustness to low- and mid-level binary variations does not generalize to high-level semantic differences, underscoring a critical blind spot in current BFSD evaluation practices.