Pith. sign in

REVIEW 4 major objections 6 minor 41 references

Pretraining on Call Graphs: When Binary Analysis Tasks Profit From Context

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Adding call-graph context to binary function embeddings improves code-similarity matching but can hurt other binary-analysis tasks.

desk verdict Useful, systematic negative result: call-graph context helps BCSD but hurts other tasks, though the claimed semantic/syntactic trade-off is under-identified. read the letter →

arxiv 2608.02084 v1 pith:LR5VFQBS submitted 2026-08-03 cs.SE cs.CRcs.LG

Samuel Valenzuela , Johannes Kinder This is my paper
classification cs.SEcs.CRcs.LG
keywords binarycodesimilaritydetectioncallgraphfunctionembeddingsneuralnetworkslabelingcompileroptimizationsemanticversussyntactictasksanalysis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that enriching binary function embeddings with call-graph context is a double-edged intervention. On binary code similarity detection (BCSD), a graph attention network with depth 2 lifts recall@1 from 0.758 to 0.867 for one backbone and from 0.579 to 0.817 for another. Yet every graph-refined embedding scores below the unrefined backbone on function labeling, and most score below on compiler-optimization detection. The paper reads this as a tension between abstractive power and technical precision: training for semantic similarity tends to weaken syntactic sensitivity. This matters because call-graph context is increasingly used to improve embeddings, and the paper shows that the choice is task-dependent, not universally beneficial.

What carries the argument

The central object is the attributed call graph: each node holds a function's pretrained embedding and each directed edge represents a caller-callee relationship. A graph neural network refines node embeddings by aggregating callee embeddings within a chosen call depth (1 to 5). The paper compares three refinements: a graph convolutional network, a graph attention network (GAT, a GNN that learns importance weights for neighboring nodes), and a simple averaging baseline. The GAT's attention mechanism does the load-bearing work, learning which callee embeddings matter, and depth 2 gives the best BCSD results. The key advantage of the mechanism is that it aggregates at function level, bypassing

What would settle it

Train a graph-refinement model on a semantic objective other than BCSD, such as function labeling, and check whether compiler-optimization detection F1 still drops; if it rises, the inverse correlation is an artifact of BCSD's training objective rather than a general trade-off. Alternatively, if a graph model that cannot exploit call-graph isomorphism still shows the same BCSD gains on hard pairs, the semantic-robustness explanation would be weakened.

Watch

Extended reading notes

Core claim

The paper's central discovery is that call-graph refinement is not a universal upgrade for binary function embeddings. Trained on BCSD with a contrastive loss, the best graph-based model substantially improves matching of same-source functions compiled with different settings, lifting recall@1 from 0.758 to 0.867 for CLAP and from 0.579 to 0.817 for jTrans. The same refined embeddings, however, consistently underperform their unrefined backbones on the semantic function-labeling task XFL, and on the syntactic compiler-optimization task COD they mostly lose as well. The authors interpret this as an inverse correlation: models that learn to abstract away compiler-induced differences become les

Load-bearing premise

The load-bearing premise is that XFL and COD cleanly stand for semantic and syntactic capability, so the observed correlations reflect a semantic-vs-syntactic divide rather than confounds from preprocessing, architecture, or downstream-head design.

Editorial extensions

If this is right

  • If the central claim is right, BCSD gains from call-graph pretraining should not be read as evidence of general semantic understanding; downstream tasks need separate evaluation.
  • Function embeddings refined for similarity search will typically be weaker for syntactic recovery tasks such as compiler-optimization detection, so one representation should not be assumed to serve both semantic and syntactic goals.
  • Call-graph context is most valuable where the base embedding is weakest: large functions, dissimilar or non-isomorphic pairs, C++ code, and extreme optimization pairs.
  • Namespace-related functions benefit more from graph refinement than functions whose identity is individual logic, suggesting the method suits context-dependent reverse-engineering scenarios.
  • The optimal context depth is shallow — about 2 for these backbones — so deeper call-graph aggregation adds noise rather than signal.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An implication left implicit: if the semantic-vs-syntactic trade-off is a general property of contrastive semantic pretraining, then any binary-analysis embedding model should be reported with at least one syntactic probe, not just BCSD or function labeling, before being called general-purpose.
  • Testable extension: training the same graph-refinement models directly on XFL or COD, instead of BCSD, would show whether the inverse correlation is caused by BCSD's objective or by call-graph context itself; the paper's setup cannot separate these.
  • The namespace finding could be probed further: an ablation that separates internal from external library callees would test whether the benefit comes from library-call patterns or from structural similarity among namespace-related functions.
  • Conditional deployment: because the gains concentrate in hard, context-heavy cases, a practical extension is on-demand refinement — apply the GNN only to functions whose base-embedding confidence or call-graph size exceeds a threshold — which the paper's grouped results suggest would preserve most of the benefit while limiting downstream loss.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper investigates whether refining binary function embeddings with call-graph context, via GNNs trained on binary code similarity detection (BCSD), benefits or harms downstream tasks. Using CLAP and jTrans as frozen backbones and AVG, GCN, and GAT aggregators at call depths 1-5, the authors report that GAT(2) substantially improves BCSD recall@1 (CLAP: 0.758 to 0.867; jTrans: 0.579 to 0.817), yet all graph-refined embeddings underperform the backbone on XFL function labeling and on compiler optimization detection (COD) for most depths. They interpret this as a trade-off between semantic similarity and syntactic precision, supported by dataset-slicing analyses showing gains are largest for functions with large call graphs, hard pairs, and namespace-related labels.

Significance. If the negative-transfer result holds, it is valuable for the binary analysis community because BCSD is widely used as a proxy for embedding quality. The paper compares two state-of-the-art backbones, releases code and data, includes a random-embedding sanity check, and its main BCSD/XFL/COD patterns in Figure 2 are visually consistent. The explanatory slicing methodology is a useful template for future work. However, the broad claim that 'optimizing for semantic similarity tasks correlates with worse performance on syntactic tasks' is not yet causally identified: the comparisons confound training objective with graph architecture and with backbone-specific preprocessing.

major comments (4)
  1. [§4.3, Fig. 2 (COD column)] The central trade-off claim is under-identified. Replacing the frozen backbone with a BCSD-trained GNN changes both the training objective (InfoNCE on BCSD) and the embedding architecture (message passing over the call graph). The AVG results show that simple aggregation can improve COD for CLAP, so the observed COD degradation is not attributable to the semantic objective alone. A control with the same GAT/GCN architecture trained on randomly shuffled labels, or an untrained GNN, is needed to separate the effect of the semantic objective from the smoothing/aggregation effect. The random-embedding sanity check in §3.1 checks implementation, not this confound. Without this control, the abstract's statement that 'optimizing for semantic similarity tasks correlates with worse performance on syntactic tasks' is not established.
  2. [Fig. 2; Tables 1-4] All results appear to come from a single training run; no error bars, confidence intervals, or significance tests are reported. Many comparative claims rely on small margins (e.g., XFL differences among AVG/GCN/GAT, or COD differences across depths). At least three seeds with variance reporting, or a statistical test, are needed for the paper's quantitative claims to be reliable.
  3. [§5, Table 4] The explanatory analysis selects the best-scoring model instance per family after seeing test results ('we focus on the best-scoring instance of each model family for the respective task'). For example, GAT(2) is used for BCSD and GAT(1) for XFL. This post hoc selection can inflate differences and makes the grouped robustness claims (e.g., namespace labels, optimization-pair robustness) difficult to interpret. The paper should either pre-specify the model instance, use a validation-based selection, or report results for all depths to show the pattern is stable.
  4. [§4.3; §6] The inverse correlation across backbones is confounded by factors other than 'semantic optimization.' CLAP and jTrans differ in preprocessing (IDA-extracted information), pretraining objectives (natural language supervision vs. jump prediction), and architecture details; the paper itself attributes CLAP's lower COD to 'reduced access to information reconstructed by IDA in the preprocessed inputs.' Therefore the CLAP-vs-jTrans comparison cannot serve as evidence for a general correlation between semantic similarity capability and syntactic task performance. The claim should be restricted to the authors' experimental setup or supported by a controlled manipulation (e.g., same backbone, different pretraining objectives).
minor comments (6)
  1. [Abstract] 'neither of semantic nor of syntactic nature' is grammatically awkward; suggest 'neither semantic nor syntactic.'
  2. [§2] The filtering description 'we only keep one function pair per source function if compilation with different optimization configurations results in binary code with differing opcode hashes' is ambiguous; clarify whether 'one pair per source function' means one pair per (source function, optimization pair) or globally.
  3. [§3.1] The hyperparameter selection is described only as 'conducting empirical evaluations to decide on fixed values'; the chosen hyperparameter values are not reported. A table or appendix listing them would improve reproducibility.
  4. [§5.4] The LLM-based label classification procedure is under-specified. State how many labels were disputed, how many were manually verified, and provide the prompt or a link to the annotation data.
  5. [Fig. 3] The x-axis label 'blog2 (Nodes)c + 1' appears mis-typeset; should be ⌊log2(Nodes)⌋ + 1.
  6. [§7] The relationship to BinEnhance, the closest framework, is described only briefly; a sentence on how the authors' setup differs experimentally (e.g., no data-co-use or string-use edges) would help position the contribution.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's central results are direct test-set measurements, not quantities reconstructed from fitted parameters or self-referential derivations.

full rationale

This is an empirical study. The main claims—that call-graph refinement improves BCSD, does not transfer to XFL, and correlates negatively with COD—are supported by directly measured test-set metrics (recall@1, micro/macro F1) on held-out function pairs. No parameter is fitted to a target quantity and then renamed a prediction: the GNNs are trained on BCSD with the InfoNCE objective and evaluated on independent downstream tasks with frozen embeddings and simple linear/tree heads. The paper even includes a random-embedding sanity check. Self-citations appear (XFL [28], BLens [4], saliency analysis [7]), but they are used as benchmarks, related work, or methodology background, not as load-bearing justifications of the observed results; XFL is an external task whose predictions are measured here rather than imported. The COD degradation is partly entailed by the BCSD training objective (positive pairs are same-source functions compiled with different optimization levels, so the contrastive objective encourages invariance to optimization), but the paper reports this as an empirical observation with comparison to backbones and acknowledges in Section 6 that the choice of training task may have restricted learned information. That is a design-interpretation limitation, not a circular derivation. No equation reduces to an input, and no fitted value is presented as a prediction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

This is an empirical ML study, so the ledger records dataset and modeling assumptions and tuned hyperparameters rather than derived constants. No new physical or architectural entities are introduced. The main uncharged inputs are the representativeness of BinaryCorp-26M, the construct validity of the semantic/syntactic task split, and the LLM-based namespace labeling scheme.

free parameters (4)
  • InfoNCE temperature = not stated
    Temperature of the NT-Xent loss is tuned on a holdout set; it affects how strongly the GNNs separate positive and negative function pairs.
  • GNN training hyperparameters (learning rate, weight decay, dropout) = not stated
    Tuned on a secluded part of the train set but values are not reported in the paper; these choices affect BCSD and downstream results.
  • Call graph depth n = 2 selected as best for GAT
    Depth is swept from 1 to 5, and the best-scoring instance per model family is selected after seeing test performance, which is a model selection choice.
  • Function filtering thresholds = min 4 instructions; one pair per source function with differing opcode hashes
    Dataset construction choices made by hand to reduce noise; they change the distribution of functions and pairs used in all experiments.
assumptions (5)
  • domain assumption CLAP and jTrans embeddings are meaningful semantic representations and cosine similarity is a valid similarity measure.
    Assumed in Section 2 and used to define the BCSD retrieval task in Section 4.1.
  • domain assumption Same-source functions compiled with different optimization levels form positive pairs and operationalize semantic equivalence.
    The entire BCSD training signal in Section 2 is built on this pairing.
  • domain assumption XFL and COD are representative semantic and syntactic downstream tasks.
    The transfer and trade-off claims in Sections 4.2 and 4.3 rest on this choice; the paper partially acknowledges the limitation in Section 6.
  • domain assumption BinaryCorp-26M with the retained train/test split supports generalizable conclusions.
    All experiments use this corpus; the paper notes in Section 6 that results may not hold for other corpora.
  • ad hoc to paper LLM-based assignment of labels to namespace versus other classes is a valid proxy for context-dependent versus individual-logic functions.
    Introduced in Section 5.4 for the label-class analysis; the classification is based on unanimous LLM opinions plus manual resolution, without an external gold standard.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pretraining on Call Graphs: When Binary Analysis Tasks Profit From Context." pith.science (2026). https://pith.science/paper/LR5VFQBS

@misc{pith2026260802084,
  author       = {Pith},
  title        = {Pith review of: Pretraining on Call Graphs: When Binary Analysis Tasks Profit From Context},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LR5VFQBS}},
  note         = {Machine review of arXiv:2608.02084}
}
read the original abstract

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.

Figures

Figures reproduced from arXiv: 2608.02084 by the authors.

Figure 1
Figure 1. Overview of our workflow. First, a binary is disassembled and all functions are extracted. These are then processed by [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Task-specific results of CLAP and jTrans (Baseline) as well as the GNNs (GCN and GAT) and averaging approach (AVG) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Grouped BCSD results of CLAP-backed models by [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Grouped BCSD results of CLAP-backed models by [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 2 canonical work pages

  1. [1]

    Hex-Rays

    Hex-Rays 2021.IDA Pro. Hex-Rays. https://www.hex-rays.com/ida-pro

    2021

  2. [2]

    Anthropic. 2025. Claude Sonnet 4. https://www.anthropic.com/news/claude-4. Accessed: 2025-08-18

    2025

  3. [3]

    Jasmijn Bastings and Katja Filippova. 2020. The elephant in the interpretability room: Why use attention as explanation when we have saliency methods?. In Proceedings of the Third BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP. Association for Computational Linguistics, Online, 149–155. doi:10.18653/v1/2020.blackboxnlp-1.14

    doi:10.18653/v1/2020.blackboxnlp-1.14 2020

  4. [4]

    Tristan Benoit, Yunru Wang, Moritz Dannehl, and Johannes Kinder. 2025. BLens: Contrastive captioning of binary functions using ensemble embedding. InPro- ceedings of the 34th USENIX Conference on Security Symposium(Seattle, WA, USA) (SEC ’25). USENIX Association, USA, Article 353, 20 pages

    2025

  5. [5]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. InPro- ceedings of the 37th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 119). PMLR, 1597–1607. https://proceedings.mlr. press/v119/chen20j.html

    2020

  6. [6]

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, Luke Marris, Sam Petulla, Colin Gaffney, Asaf Aharoni, Nathan Lintz, Tiago Pais, Henrik Jacobsson, Idan Szpektor, Nan-Jiang Jiang, and Chris Hahn. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning...

    arXiv 2025

  7. [7]

    Moritz Dannehl, Samuel Valenzuela, and Johannes Kinder. 2025. Which in- structions matter the most: A saliency analysis of binary function embed- ding models. In2025 IEEE Security and Privacy Workshops (SPW). 145–151. doi:10.1109/SPW67851.2025.00019

    arXiv 2025

  8. [8]

    Yaniv David, Uri Alon, and Eran Yahav. 2020. Neural reverse engineering of stripped binaries using augmented control flow graphs.Proceedings of the ACM on Programming Languages4, OOPSLA, Article 225 (Nov. 2020), 28 pages. doi:10. 1145/3428293

    2020

Show all 41 references
  1. [9]

    DeepSeek AI. 2025. DeepSeek-R1-0528. https://huggingface.co/deepseek-ai/ DeepSeek-R1-0528. Accessed: 2025-09-20

    2025

  2. [10]

    Steven H. H. Ding, Benjamin C. M. Fung, and Philippe Charland. 2019. Asm2Vec: Boosting static representation robustness for binary clone search against code obfuscation and compiler optimization. In2019 IEEE Symposium on Security and Privacy (SP). 472–489. doi:10.1109/SP.2019.00003

    2019

  3. [11]

    Jian Gao, Xin Yang, Ying Fu, Yu Jiang, and Jiaguang Sun. 2018. VulSeeker: A semantic learning based vulnerability seeker for cross-platform binary. In2018 33rd IEEE/ACM International Conference on Automated Software Engineering (ASE). 896–899. doi:10.1145/3238147.3240480

    2018

  4. [12]

    Yixin Guo, Pengcheng Li, Yingwei Luo, Xiaolin Wang, and Zhenlin Wang. 2022. Exploring GNN based program embedding technologies for binary related tasks. InProceedings of the 30th IEEE/ACM International Conference on Program Compre- hension(Virtual Event)(ICPC ’22). Association...

    2022

  5. [13]

    Aric A Hagberg, Daniel A Schult, and Pieter J Swart. 2008. Exploring network structure, dynamics, and function using NetworkX. InProceedings of the 7th Python in Science Conference. SciPy, Pasadena, CA, 11–15

    2008

  6. [14]

    Jingxuan He, Pesho Ivanov, Petar Tsankov, Veselin Raychev, and Martin Vechev

  7. [15]

    Xu He, Shu Wang, Yunlong Xing, Pengbin Feng, Haining Wang, Qi Li, Songqing Chen, and Kun Sun. 2022. BinProv: Binary code provenance identification without disassembly. InProceedings of the 25th International Symposium on Research in Attacks, Intrusions and Defenses(Limassol, C...

    2022 doi

  8. [16]

    Howie Huang

    Yuede Ji, Lei Cui, and H. Howie Huang. 2021. Vestige: Identifying binary code provenance for vulnerability detection. InApplied Cryptography and Network Security: 19th International Conference, ACNS 2021, Kamakura, Japan, June 21–24, 2021, Proceedings, Part II(Kamakura, Japan)...

    2021 doi

  9. [17]

    Xin Jin, Kexin Pei, Jun Yeon Won, and Zhiqiang Lin. 2022. SymLM: Predicting function names in stripped binaries via context-sensitive execution-aware code embeddings. InProceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security(Los Angeles, CA, USA)(...

    2022 doi

  10. [19]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. InProceedings of the 5th International Conference on Learning Representations(Palais des Congrès Neptune, Toulon, France)(ICLR ’17)

    2017

  11. [20]

    Xuezixiang Li, Yu Qu, and Heng Yin. 2021. PalmTree: Learning an assembly language model for instruction embedding. InProceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security(Virtual Event, Republic of Korea)(CCS ’21). Association for Computing Mach...

    2021

  12. [21]

    Bingchang Liu, Wei Huo, Chao Zhang, Wenchao Li, Feng Li, Aihua Piao, and Wei Zou. 2018. 𝛼Diff: Cross-version binary code similarity detection with DNN. InProceedings of the 33rd ACM/IEEE International Conference on Automated Soft- ware Engineering(Montpellier, France)(ASE ’18)...

    2018

  13. [22]

    Zhenhao Luo, Pengfei Wang, Baosheng Wang, Yong Tang, Wei Xie, Xu Zhou, Danjun Liu, and Kai Lu. 2023. VulHawk: Cross-architecture vulnerability detection with entropy-based binary code search. In30th Annual Network and Distributed System Security Symposium(San Diego, CA, USA)(N...

    2023

  14. [23]

    Luca Massarelli, Giuseppe Antonio Di Luna, Fabio Petroni, Roberto Baldoni, and Leonardo Querzoni. 2019. SAFE: Self-attentive function embeddings for binary similarity. InDetection of Intrusions and Malware, and Vulnerability Assessment. Springer International Publishing, Cham, 309–329

    2019

  15. [24]

    Luca Massarelli, Giuseppe Antonio Di Luna, Fabio Petroni, Leonardo Querzoni, Roberto Baldoni, et al. 2019. Investigating graph embedding neural networks with unsupervised features extraction for binary analysis. InProceedings BAR 2019 Workshop on Binary Analysis Research. 1–11...

    2019

  16. [25]

    Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space.arXiv preprint arXiv:1301.3781 (2013)

    2013 arXiv

  17. [26]

    Ziad Obermeyer, Brian Powers, Christine Vogeli, and Sendhil Mullainathan. 2019. Dissecting racial bias in an algorithm used to manage the health of populations. Science366, 6464 (2019), 447–453. doi:10.1126/science.aax2342

    2019 doi

  18. [27]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748(2018)

    2018 arXiv

  19. [28]

    James Patrick-Evans, Moritz Dannehl, and Johannes Kinder. 2023. XFL: Naming functions in binaries with extreme multi-label learning. In2023 IEEE Sympo- sium on Security and Privacy (S&P). IEEE, 2375–2390. doi:10.1109/SP46215.2023. 10179439

    2023

  20. [29]

    Kexin Pei, Zhou Xuan, Junfeng Yang, Suman Jana, and Baishakhi Ray. 2023. Learning approximate execution semantics from traces for binary function similarity.IEEE Transactions on Software Engineering49, 4 (2023), 2776–2790. doi:10.1109/TSE.2022.3231621

    2023

  21. [30]

    Nicholas

    Edward Raff, Jon Barker, Jared Sylvester, Robert Brandon, Bryan Catanzaro, and Charles K. Nicholas. 2018. Malware detection by eating a whole EXE. InThe Workshops of the Thirty-Second AAAI Conference on Artificial Intelligence(New Orleans, LA, USA)(AAAI Technical Report, Vol. ...

    2018

  22. [31]

    Miller, and Xiaojin Zhu

    Nathan Rosenblum, Barton P. Miller, and Xiaojin Zhu. 2011. Recovering the toolchain provenance of binary code. InProceedings of the 2011 International Symposium on Software Testing and Analysis(Toronto, Ontario, Canada)(ISSTA ’11). Association for Computing Machinery, New York...

    2011

  23. [32]

    Thalles Santos Silva. 2020. Exploring SimCLR: A simple framework for contrastive learning of visual representations.https://sthalles.github.io(2020). https://sthalles. github.io/simple-self-supervised-learning/

    2020

  24. [33]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. InProceedings of the 31st International Conference on Neural Informa- tion Processing Systems(Long Beach, CA, USA)(NIP...

    2017

  25. [34]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lió, and Yoshua Bengio. 2018. Graph attention networks. InProceedings of the 6th International Conference on Learning Representations(Vancouver, BC, Canada) (ICLR ’18)

    2018

  26. [35]

    Hao Wang, Zeyu Gao, Chao Zhang, Zihan Sha, Mingyang Sun, Yuchen Zhou, Wenyu Zhu, Wenju Sun, Han Qiu, and Xi Xiao. 2024. CLAP: Learning transferable binary code representations with natural language supervision. InProceedings of 11 ICPC ’26, April 12–13, 2026, Rio de Janeiro, B...

    2024

  27. [36]

    Hao Wang, Wenjie Qu, Gilad Katz, Wenyu Zhu, Zeyu Gao, Han Qiu, Jianwei Zhuge, and Chao Zhang. 2022. jTrans: Jump-aware transformer for binary code similarity detection. InProceedings of the 31st ACM SIGSOFT International Symposium on Software Testing and Analysis(Virtual, Sout...

    2022

  28. [37]

    Yongpan Wang, Hong Li, Xiaojie Zhu, Siyuan Li, Chaopeng Dong, Shouguo Yang, and Kangyuan Qin. 2025. BinEnhance: An enhancement framework based on external environment semantics for binary code search. In32nd Annual Network and Distributed System Security Symposium(San Diego, C...

    2025

  29. [38]

    Xiangzhe Xu, Shiwei Feng, Yapeng Ye, Guangyu Shen, Zian Su, Siyuan Cheng, Guanhong Tao, Qingkai Shi, Zhuo Zhang, and Xiangyu Zhang. 2023. Improv- ing binary code similarity transformer models by semantics-driven instruction deemphasis. InProceedings of the 32nd ACM SIGSOFT Int...

    2023

  30. [39]

    Xiaojun Xu, Chang Liu, Qian Feng, Heng Yin, Le Song, and Dawn Song. 2017. Neural network-based graph embedding for cross-platform binary code similarity detection. InProceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security(Dallas, TX, USA)(CCS ’17)...

    2017

  31. [40]

    Shih-Yuan Yu, Yonatan Gizachew Achamyeleh, Chonghan Wang, Anton Ko- cheturov, Patrick Eisen, and Mohammad Abdullah Al Faruque. 2023. CFG2VEC: Hierarchical graph neural network for cross-architectural software reverse engi- neering. InProceedings of the 45th International Confe...

    2023

  32. [41]

    Zeping Yu, Rui Cao, Qiyi Tang, Sen Nie, Junzhou Huang, and Shi Wu. 2020. Order Matters: Semantic-aware neural networks for binary code similarity detection. Proceedings of the AAAI Conference on Artificial Intelligence34, 01 (Apr. 2020), 1145–1152. doi:10.1609/aaai.v34i01.5466 12

    2020 doi

  33. [2018]

    InProceedings of the 2018 ACM SIGSAC Conference on Computer and Communications Security (Toronto, Canada)(CCS ’18)

    Debin: Predicting debug information in stripped binaries. InProceedings of the 2018 ACM SIGSAC Conference on Computer and Communications Security (Toronto, Canada)(CCS ’18). Association for Computing Machinery, New York, NY, USA, 1667–1680. doi:10.1145/3243734.3243866

    2018

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.