Pith. sign in

REVIEW 6 major objections 5 minor 61 references

KGBERT4Eth: A Feature-Complete Transformer Powered by Knowledge Graph for Multi-Task Ethereum Fraud Detection

T0 review · 6 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read KGBERT4Eth claims that jointly pre-training a transaction language model and a transaction knowledge graph yields state-of-the-art Ethereum fraud detection, with absolute F1 gains of 8–16 points on phishing detection and 6–26 points on acco

desk verdict Solid, incremental pre-training paper for Ethereum fraud detection with big F1 numbers, but the 'feature-complete' claim leans on a graph branch that is barely present at inference and whose formal description has a hole. read the letter →

arxiv 2509.03860 v1 pith:SOGTGBCD submitted 2025-09-04 cs.CR

Yifan Jia , Ye Tian , Liguo Zhang , Yanbin Wang , Jianguo Sun , Liangliang Song This is my paper
classification cs.CR
keywords Ethereumfrauddetectionphishingde-anonymizationtransactionknowledgegraphpre-trainedtransformermaskedlanguagemodelinglinkpredictionBM25
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 proposes KGBERT4Eth, a pre-trained encoder that fuses two views of Ethereum transaction data: a language model reading textualized transaction records and a knowledge graph encoding fund flows plus expert-designed account features. The two components are pre-trained jointly, with biased masked token prediction on the text side and link prediction on the graph side, coordinated by a mask-invariant attention mechanism. The authors report that this feature-complete representation consistently beats 15 baselines, with absolute F1 improvements of 8–16 percentage points across three phishing detection benchmarks and 6–26 points across four de-anonymization datasets. If correct, the result would mean that sequential transaction semantics and graph-structured fund-flow knowledge can be fused during pre-training and then used at inference without explicitly constructing the graph.

What carries the argument

The load-bearing mechanism is the mask-invariant attention synergy (MiAS): a cross-attention module, inspired by prior knowledge-enhanced language modeling [44], that uses a visibility matrix so only the [CLS] token of a transaction sequence may attend to that account's knowledge-graph entities, while all other token positions are blocked from receiving graph information. This preserves the masked-token prediction objective while still letting structured knowledge flow into the account-level embedding. Around it, the pre-training loss combines Biased Mask Prediction (BMP), which uses BM25 scores to mask statistically rare and discriminative tokens, and Transaction Link Prediction (TLP), whic

What would settle it

A probing experiment would settle this: after fine-tuning, train a linear probe to predict expert graph features (e.g., account lifespan, Katz centrality, or degree centrality) from the [CLS] embedding; if the probe performs at chance, the graph-derived knowledge is not actually retained in the representation used for downstream tasks. Alternatively, an ablation that blocks the TKG branch entirely during pre-training and compares phishing F1 would quantify how much of the reported 87–88 F1 truly depends on graph transfer.

Watch

Extended reading notes

Core claim

The central claim is that Ethereum fraud detection improves when one encoder is pre-trained to model transaction text and transaction-graph structure together, rather than forcing practitioners to choose among expert features, graph embeddings, or sequential patterns. The paper argues that joint optimization of a biased mask-prediction objective and a transaction link-prediction objective produces 'feature-complete' account embeddings, and that the proposed mask-invariant attention synergy lets graph knowledge enrich language representations without corrupting the masked-token learning signal. On three phishing detection datasets the model reaches F1 scores of 87.60, 85.43, and 88.02, and on

Load-bearing premise

The reported gains assume that graph-derived structural and expert knowledge learned during pre-training transfers to completely unseen accounts through the [CLS] representation, even though the knowledge graph is never built or updated when the model is fine-tuned and evaluated.

Editorial extensions

If this is right

  • Across three phishing benchmarks, the model's F1 exceeds the strongest baseline by 8–16 absolute points, with false-negative rates roughly halved in several settings.
  • On de-anonymization, the model improves overall F1 by 6–26 absolute points, with particularly large gains on sparse identity classes such as ICO wallets.
  • At inference the model needs only textual transaction records; the knowledge graph is not reconstructed or updated, so deployment cost stays close to a text encoder.
  • Ablation results indicate each proposed component contributes: removing biased masking, the knowledge graph, expert features, or the synergy module each lowers F1, and generic pretrained encoders underperform all variants.

Reading between the lines

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

  • The paper's inductive evaluation—pre-training accounts excluded from downstream—is a strong setup, but the absence of a graph at fine-tuning means the KG's value must be carried entirely by fixed [CLS] representations; a probe checking whether those representations encode graph-derived node features would make that mechanism visible.
  • The same two-stage recipe—textualize structured records, pre-train with a knowledge graph, discard the graph at inference—could transfer to other blockchains or structured audit logs where records are homogeneous and labels are rare, though the paper does not test this.
  • Because the BM25 masking threshold and the two-task loss steer the model toward rare anomalous tokens, performance is likely sensitive to corpus composition; on a corpus with different rarity statistics the threshold would need recalibration.
  • The experiments do not fully isolate whether the gains come from the expert node features, the graph topology, or the mask-invariant fusion; the ablations show each matters but not the causal path.
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

6 major / 5 minor

Summary. The paper proposes KGBERT4Eth, a pre-training encoder for Ethereum fraud detection that combines a Transformer language model over textualized transaction records with a Transaction Knowledge Graph whose node features include expert-curated statistical, temporal, and structural features. The training procedure uses BM25-based biased masking, knowledge-graph link prediction, and a mask-invariant cross-attention module to fuse the two modalities. The authors evaluate the model on two downstream tasks—phishing account detection on three datasets and account de-anonymization on four account-type datasets—and report large F1 improvements over fifteen baselines, together with ablations and hyperparameter sensitivity analyses.

Significance. If the reported gains hold, the paper would make a useful contribution to Ethereum fraud detection by demonstrating that sequential and graph/expert knowledge can be integrated during pre-training. The manuscript is generally well structured, includes public benchmark datasets, reports standard errors, provides ablation studies and a sensitivity analysis, and makes source code available. However, the evaluation protocol leaves a key question open: at test time the knowledge graph is not built or updated, so the claimed advantage of cross-paradigm fusion depends on whether graph-derived knowledge survives in the [CLS] representation. The current evidence suggests that the knowledge-graph component contributes only a small part of the improvement, and several specification gaps weaken the support for the headline claims.

major comments (6)
  1. The evaluation setup is inconsistent with the central architectural claim. §VI-A states that during fine-tuning and evaluation the model processes textual records 'without building or updating the KG' and that all accounts used in pre-training are excluded from downstream evaluation. Consequently, test accounts have no node in the TKG and the GNN branch cannot process them; graph/expert knowledge can affect predictions only through the [CLS] representation learned during pre-training. No experiment or analysis shows that [CLS] retains such knowledge for unseen accounts. The w/o TKG ablation in Table V removes the KG and changes F1 by only -1.1/-1.6 points, suggesting the KG contributes a small part of the reported gain. To support the 'feature-complete cross-paradigm' claim, the authors should either build the KG for downstream accounts at fine-tuning/evaluation, or provide direct eviden
  2. The MiAS update is formally under-specified. In Eq. (10), the visibility matrix M is said to 'set the attention scores of all elements in \tilde{h}^{(l)} except \tilde{h}^{(l)}_{CLS} to 0'. For non-CLS query positions, this makes the attention weights α_{h,v} zero, so \hat{h}^{(l)} would be the zero vector unless a residual connection is implicitly assumed. The text claims that sentence semantics are preserved, but no residual term is written. Please state the exact form of M (e.g., -inf for invisible positions) and the complete update with any residual/projection.
  3. The paper does not compare against the closest prior work. TLMG4Eth [10] and the dynamic feature fusion method [11] are cited in the introduction as the most relevant joint language/graph models for Ethereum fraud detection, but neither appears in Tables III/IV or in the baseline descriptions of Section VI-C. Since KGBERT4Eth is explicitly positioned as an enhancement of this line of work, the absence of these baselines makes the claim of 'significantly outperforms state-of-the-art' incomplete.
  4. The abstract's reported improvement range for de-anonymization (6–26%) is inconsistent with Table IV. For ICO Wallets, the proposed method reaches F1=65.66 vs. the best baseline TSGN at 36.10, an absolute gain of 29.56 points. Even if the claim refers to overall F1 (90.52 vs. 81.07 = 9.45), the range should be stated per-dataset and reconciled. Please correct the abstract or clarify which comparison is meant.
  5. The ablation results are internally surprising and not discussed: w/o Expert (84.49/88.22) is worse than w/o TKG (86.95/88.91) on both tasks. That is, removing the entire KG (including structural features) improves over keeping the KG with randomly initialized entity features. This suggests the structural branch alone is harmful and the expert features are what matter; the paper's conclusion that 'each component is integral' does not follow. Please explain this interaction or provide additional ablations (e.g., expert features without KG, or structural features with expert initialization).
  6. The experimental setup is underspecified for verification. It is not reported how baselines were tuned, what hyperparameter search was used, which train/validation splits were used for each baseline, or how the 'Overall' de-anonymization metric is computed. The sentence 'Finetuning datasets employed 2-hop node retrieval with a maximum of 100 nodes' is also unclear given that no KG is built during fine-tuning. Without these details, the large absolute differences in Tables III/IV cannot be fully assessed.
minor comments (5)
  1. There are typographical errors: 'a enhanced' in the abstract, 'B ACKGROUD' in Section II, 'TML' vs. 'TLM' in the Introduction, 'invoked invoked' in Table I, and 'Mev Bot' should be 'MEV Bot'. These should be corrected before publication.
  2. The 'feature-complete' claim is overstated because the TLM input keeps only five of the nineteen transaction fields listed in Table I. The exclusion of fields such as isError, txreceipt_status, functionName, and methodId is asserted without justification; an ablation on the retained field set would be helpful.
  3. The expert features were designed after an empirical analysis of phishing account statistics (Fig. 1). This label-informed feature selection is a potential source of optimism, even though the final evaluation is on held-out accounts. The authors should acknowledge this and, if possible, validate the feature design on data whose labels were not used in feature selection.
  4. No statistical significance tests are reported beyond standard errors. Given that some comparisons (e.g., the score functions in Table VI) differ by less than a point, a paired significance test would strengthen the conclusions.
  5. The sensitivity plots would be clearer with error bars or confidence bands on the F1 values, especially since all main tables report standard errors.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: downstream evaluation is held-out against external baselines; the KG-absence and label-informed feature concerns are limitations or correctness risks, not circular derivations.

full rationale

The core performance claim is tested on downstream accounts that were excluded from pre-training, and the model is compared against external baselines (BERT4ETH, ZipZap, GrabPhisher, etc.), so the reported F1 gains are not forced by construction. The paper explicitly states that the KG is not built or updated during fine-tuning/evaluation, which makes the transfer of graph knowledge through the [CLS] representation an assumption rather than a circular identity; the w/o TKG ablation (Table V) provides an empirical, albeit modest, test of the KG's contribution. The expert features in Table II are motivated by the empirical phishing-account distributions in Fig. 1, which is a potential label-informed feature-selection/leakage concern, but those features are computed from transaction metadata and graph topology and do not encode the target label, so they do not make the prediction equal to the input by definition. The self-citation [10] is used as background and as the source of the SPN dataset, but it is not load-bearing for the central claim and no uniqueness theorem or ansatz is imported from the authors' prior work to force the model choice. The MiAS update in Eq. 10 is technically underspecified, but that is an ambiguity, not a circular reduction. Overall, the derivation chain is self-contained enough that no prediction reduces to its own input by construction.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim depends on a large set of hand-chosen hyperparameters and features. The model introduces no new physical or mathematical entities; the TKG is a constructed graph used only during pre-training. The most load-bearing domain assumptions are that BM25 rarity signals fraud-relevant informativeness, that two edge types suffice for fund-flow patterns, and that structural knowledge transfers to accounts absent from the TKG at test time.

free parameters (5)
  • mask threshold τ = 0.1
    Determines which tokens are masked in Biased Mask Prediction; sensitivity analysis (Fig. 3) shows a narrow optimal range and sharp degradation for τ > 0.3.
  • maximum input length L_max = 512
    Truncation length for the BERT encoder; Fig. 4 shows performance consistently improves with L_max, so the chosen value is an upper bound imposed by the architecture rather than an optimum.
  • BM25 parameters b and z1 = BM25L defaults (not stated numerically)
    Control token-frequency and length normalization in the BM25 score; the paper states BM25L is used but does not report values.
  • GNN depth and hidden dimension = not reported
    Number of layers and hidden size for the TKG encoder are not specified, yet they determine the structural representation capacity.
  • training hyperparameters (learning rate, batch size, epochs) = not reported
    Critical for reproducing the reported F1 values but absent from the paper.
assumptions (6)
  • standard math BM25 formula (Equation 2) measures token informativeness
    Standard IR ranking function, used as the basis for biased masking.
  • domain assumption Transformer encoder and BERT tokenizer are appropriate for textualized transactions
    The paper assumes that converting numeric transaction fields to strings preserves semantics useful for fraud detection (Section V-A).
  • domain assumption Two relation types in the TKG (external vs contract transactions) capture fund-flow patterns
    The knowledge graph reduces all Ethereum interactions to two edge types, which may discard contract-specific behavior such as method calls.
  • domain assumption Expert features (Table II) are predictive of fraud and deanonymization
    Features were selected after analyzing the distribution of phishing accounts (Fig. 1), so their predictive power is assumed rather than derived.
  • domain assumption Downstream datasets and Etherscan labels are accurate ground truth
    Phishing and identity labels are community-sourced and may contain noise.
  • domain assumption Pre-training and downstream account sets are disjoint
    The paper states exclusion, but the 2-hop retrieval during fine-tuning may introduce pre-training accounts as neighborhood context.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KGBERT4Eth: A Feature-Complete Transformer Powered by Knowledge Graph for Multi-Task Ethereum Fraud Detection." pith.science (2026). https://pith.science/paper/SOGTGBCD

@misc{pith2026250903860,
  author       = {Pith},
  title        = {Pith review of: KGBERT4Eth: A Feature-Complete Transformer Powered by Knowledge Graph for Multi-Task Ethereum Fraud Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SOGTGBCD}},
  note         = {Machine review of arXiv:2509.03860}
}
read the original abstract

Ethereum's rapid ecosystem expansion and transaction anonymity have triggered a surge in malicious activity. Detection mechanisms currently bifurcate into three technical strands: expert-defined features, graph embeddings, and sequential transaction patterns, collectively spanning the complete feature sets of Ethereum's native data layer. Yet the absence of cross-paradigm integration mechanisms forces practitioners to choose between sacrificing sequential context awareness, structured fund-flow patterns, or human-curated feature insights in their solutions. To bridge this gap, we propose KGBERT4Eth, a feature-complete pre-training encoder that synergistically combines two key components: (1) a Transaction Semantic Extractor, where we train an enhanced Transaction Language Model (TLM) to learn contextual semantic representations from conceptualized transaction records, and (2) a Transaction Knowledge Graph (TKG) that incorporates expert-curated domain knowledge into graph node embeddings to capture fund flow patterns and human-curated feature insights. We jointly optimize pre-training objectives for both components to fuse these complementary features, generating feature-complete embeddings. To emphasize rare anomalous transactions, we design a biased masking prediction task for TLM to focus on statistical outliers, while the Transaction TKG employs link prediction to learn latent transaction relationships and aggregate knowledge. Furthermore, we propose a mask-invariant attention coordination module to ensure stable dynamic information exchange between TLM and TKG during pre-training. KGBERT4Eth significantly outperforms state-of-the-art baselines in both phishing account detection and de-anonymization tasks, achieving absolute F1-score improvements of 8-16% on three phishing detection benchmarks and 6-26% on four de-anonymization datasets.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 40 canonical work pages

  1. [10]

    Ethereum fraud detection via joint transaction language model and graph representation learning,

    J. Sun, Y . Jia, Y . Wang, Y . Tian, and S. Zhang, “Ethereum fraud detection via joint transaction language model and graph representation learning,” Information Fusion, vol. 120, p. 103074, 2025

    work page 2025

  2. [11]

    Dynamic feature fusion: Combining global graph structures and local semantics for blockchain phishing detection,

    Z. Sheng, L. Song, and Y . Wang, “Dynamic feature fusion: Combining global graph structures and local semantics for blockchain phishing detection,” IEEE Transactions on Network and Service Management , 2025

    work page 2025

  3. [1]

    Catalog: Exploiting joint temporal dependencies for enhanced phishing detection on ethereum,

    M. Ghosh, S. Srivastava, A. Upadhyaya, R. Halder, and J. Chandra, “Catalog: Exploiting joint temporal dependencies for enhanced phishing detection on ethereum,” in Proceedings of the ACM on Web Conference 2025, 2025, pp. 969–977

    2025

  4. [2]

    Blockchain cross-chain bridge security: Challenges, solutions, and fu- ture outlook,

    N. Li, M. Qi, Z. Xu, X. Zhu, W. Zhou, S. Wen, and Y . Xiang, “Blockchain cross-chain bridge security: Challenges, solutions, and fu- ture outlook,” Distributed Ledger Technologies: Research and Practice , vol. 4, no. 1, pp. 1–34, 2025

    work page 2025

  5. [3]

    Dapphunter: Identifying inconsistent behaviors of blockchain-based decentralized applications,

    J. Zhou, T. Jiang, H. Wang, M. Wu, and T. Chen, “Dapphunter: Identifying inconsistent behaviors of blockchain-based decentralized applications,” in 2023 IEEE/ACM 45th International Conference on Software Engineering: Software Engineering in Practice (ICSE-SEIP) . IEEE, 2023, pp. 24–35

    work page 2023

  6. [4]

    Dissecting Payload-based Transaction Phishing on Ethereum

    Z. Chen, Y . Hu, B. He, D. Luo, L. Wu, and Y . Zhou, “Dissect- ing payload-based transaction phishing on ethereum,” arXiv preprint arXiv:2409.02386, 2024

    work page Pith review arXiv 2024

  7. [5]

    Toward understanding asset flows in crypto money laundering through the lenses of ethereum heists,

    J. Wu, D. Lin, Q. Fu, S. Yang, T. Chen, Z. Zheng, and B. Song, “Toward understanding asset flows in crypto money laundering through the lenses of ethereum heists,” IEEE Transactions on Information Forensics and Security, vol. 19, pp. 1994–2009, 2023

    work page 1994

  8. [6]

    Measuring and relieving the over-smoothing problem for graph neural networks from the topological view,

    D. Chen, Y . Lin, W. Li, P. Li, J. Zhou, and X. Sun, “Measuring and relieving the over-smoothing problem for graph neural networks from the topological view,” inProceedings of the AAAI conference on artificial intelligence, vol. 34, no. 04, 2020, pp. 3438–3445

    work page 2020

Show all 61 references
  1. [7]

    Not too little, not too much: a theoretical analysis of graph (over) smoothing,

    N. Keriven, “Not too little, not too much: a theoretical analysis of graph (over) smoothing,” Advances in Neural Information Processing Systems, vol. 35, pp. 2268–2281, 2022

    2022

  2. [8]

    Bert4eth: A pre- trained transformer for ethereum fraud detection,

    S. Hu, Z. Zhang, B. Luo, S. Lu, B. He, and L. Liu, “Bert4eth: A pre- trained transformer for ethereum fraud detection,” in Proceedings of the ACM Web Conference 2023 , 2023, pp. 2189–2197

    2023

  3. [9]

    Zipzap: Efficient training of language models for large-scale fraud detection on blockchain,

    S. Hu, T. Huang, K.-H. Chow, W. Wei, Y . Wu, and L. Liu, “Zipzap: Efficient training of language models for large-scale fraud detection on blockchain,” in Proceedings of the ACM Web Conference 2024 , 2024, pp. 2807–2816

    2024

  4. [12]

    Detecting ponzi schemes on ethereum: Towards healthier blockchain technology,

    W. Chen, Z. Zheng, J. Cui, E. Ngai, P. Zheng, and Y . Zhou, “Detecting ponzi schemes on ethereum: Towards healthier blockchain technology,” in Proceedings of the 2018 world wide web conference, 2018, pp. 1409– 1418

    2018

  5. [13]

    Understanding blockchain: Definitions, architecture, design, and system comparison,

    M. H. Tabatabaei, R. Vitenberg, and N. R. Veeraragavan, “Understanding blockchain: Definitions, architecture, design, and system comparison,” Computer Science Review , vol. 50, p. 100575, 2023

    2023

  6. [14]

    Ai-powered fraud detection in decentralized finance: A project life cycle perspective,

    B. Luo, Z. Zhang, Q. Wang, A. Ke, S. Lu, and B. He, “Ai-powered fraud detection in decentralized finance: A project life cycle perspective,”ACM Computing Surveys, vol. 57, no. 4, pp. 1–38, 2024

    2024

  7. [15]

    Who stole my nft? investigating web3 nft phishing scams on ethereum,

    J. Yang, J. Liu, D. Lin, J. Wu, B. Huang, Q. Li, and Z. Zheng, “Who stole my nft? investigating web3 nft phishing scams on ethereum,” IEEE Transactions on Information Forensics and Security , 2024

    2024

  8. [16]

    Defort: Automatic detection and analysis of price manipulation attacks in defi applications,

    M. Xie, M. Hu, Z. Kong, C. Zhang, Y . Feng, H. Wang, Y . Xue, H. Zhang, Y . Liu, and Y . Liu, “Defort: Automatic detection and analysis of price manipulation attacks in defi applications,” in Proceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and A...

    2024

  9. [18]

    Txphishscope: Towards detecting and understanding transaction-based phishing on ethereum,

    B. He, Y . Chen, Z. Chen, X. Hu, Y . Hu, L. Wu, R. Chang, H. Wang, and Y . Zhou, “Txphishscope: Towards detecting and understanding transaction-based phishing on ethereum,” in Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security , 2023, pp. 120–134

    2023

  10. [19]

    Characterizing ethereum address poisoning attack,

    S. Guan and K. Li, “Characterizing ethereum address poisoning attack,” in Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security , 2024, pp. 986–1000

    2024

  11. [20]

    A novel hybrid feature fusion model for detecting phishing scam on ethereum using deep neural network,

    T. Wen, Y . Xiao, A. Wang, and H. Wang, “A novel hybrid feature fusion model for detecting phishing scam on ethereum using deep neural network,” Expert Systems with Applications , vol. 211, p. 118463, 2023

    2023

  12. [21]

    Transaction-based classification and detection approach for ethereum smart contract,

    T. Hu, X. Liu, T. Chen, X. Zhang, X. Huang, W. Niu, J. Lu, K. Zhou, and Y . Liu, “Transaction-based classification and detection approach for ethereum smart contract,” Information Processing & Management , vol. 58, no. 2, p. 102462, 2021

    2021

  13. [22]

    Enhanc- ing ethereum fraud detection via generative and contrastive self- supervision,

    C. Jin, J. Zhou, C. Xie, S. Yu, Q. Xuan, and X. Yang, “Enhanc- ing ethereum fraud detection via generative and contrastive self- supervision,” IEEE Transactions on Information Forensics and Security, 2024

    2024

  14. [23]

    2dynethnet: A two-dimensional streaming framework for ethereum phishing scam detection,

    J. Yang, W. Yu, J. Wu, D. Lin, Z. Wu, and Z. Zheng, “2dynethnet: A two-dimensional streaming framework for ethereum phishing scam detection,” IEEE Transactions on Information Forensics and Security , 2024

    2024

  15. [24]

    Towards effective detection of ponzi schemes on ethereum with contract runtime behavior graph,

    R. Liang, J. Chen, C. Wu, K. He, Y . Wu, W. Sun, R. Du, Q. Zhao, and Y . Liu, “Towards effective detection of ponzi schemes on ethereum with contract runtime behavior graph,” ACM Transactions on Software Engineering and Methodology , vol. 34, no. 4, pp. 1–32, 2025

    2025

  16. [25]

    Treat: Temporal and relational attention-based tensor representation learning for ethereum phishing users,

    M. Ghosh, R. Halder, and J. Chandra, “Treat: Temporal and relational attention-based tensor representation learning for ethereum phishing users,” IEEE Transactions on Services Computing , 2025

    2025

  17. [26]

    Phishing scam detec- tion on ethereum: Towards financial security for blockchain ecosystem

    W. Chen, X. Guo, Z. Chen, Z. Zheng, and Y . Lu, “Phishing scam detec- tion on ethereum: Towards financial security for blockchain ecosystem.” in IJCAI, vol. 7, 2020, pp. 4456–4462

    2020

  18. [27]

    Ewdps: A novel framework for early warning and detection on ethereum phishing scams,

    C. Xu, R. Li, L. Zhu, X. Shen, and K. Sharif, “Ewdps: A novel framework for early warning and detection on ethereum phishing scams,” IEEE Internet of Things Journal , 2024

    2024

  19. [28]

    Behavior- aware account de-anonymization on ethereum interaction graph,

    J. Zhou, C. Hu, J. Chi, J. Wu, M. Shen, and Q. Xuan, “Behavior- aware account de-anonymization on ethereum interaction graph,” IEEE Transactions on Information Forensics and Security , vol. 17, pp. 3433– 3448, 2022

    2022

  20. [29]

    Phishing detection on ethereum via attributed ego-graph embedding,

    Y . Xia, J. Liu, and J. Wu, “Phishing detection on ethereum via attributed ego-graph embedding,” IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 69, no. 5, pp. 2538–2542, 2022

    2022

  21. [30]

    Fishing for fraudsters: Uncovering ethereum phishing gangs with blockchain data,

    J. Liu, J. Chen, J. Wu, Z. Wu, J. Fang, and Z. Zheng, “Fishing for fraudsters: Uncovering ethereum phishing gangs with blockchain data,” IEEE Transactions on Information Forensics and Security , vol. 19, pp. 3038–3050, 2024

    2024

  22. [31]

    Phishing scams detection in ethereum transaction network,

    L. Chen, J. Peng, Y . Liu, J. Li, F. Xie, and Z. Zheng, “Phishing scams detection in ethereum transaction network,” ACM Transactions on Internet Technology (TOIT) , vol. 21, no. 1, pp. 1–16, 2020

    2020

  23. [32]

    Who are the phishers? phishing scam detection on ethereum via network embedding,

    J. Wu, Q. Yuan, D. Lin, W. You, W. Chen, C. Chen, and Z. Zheng, “Who are the phishers? phishing scam detection on ethereum via network embedding,” IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 52, no. 2, pp. 1156–1166, 2020

    2020

  24. [33]

    Deepwalk: Online learning of social representations,

    B. Perozzi, R. Al-Rfou, and S. Skiena, “Deepwalk: Online learning of social representations,” in Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining , 2014, pp. 701–710

    2014

  25. [34]

    Phish2vec: A temporal and heterogeneous network embedding approach for detecting phishing scams on ethereum,

    Z. Lin, X. Xiao, G. Hu, B. Zhang, Q. Liu, and X. Luo, “Phish2vec: A temporal and heterogeneous network embedding approach for detecting phishing scams on ethereum,” in 2023 20th Annual IEEE International Conference on Sensing, Communication, and Networking (SECON) . IEEE, 2023...

    2023

  26. [35]

    Phishing fraud detection on ethereum using graph neural network,

    P. Li, Y . Xie, X. Xu, J. Zhou, and Q. Xuan, “Phishing fraud detection on ethereum using graph neural network,” in International Conference on Blockchain and Trustworthy Systems . Springer, 2022, pp. 362–375

    2022

  27. [36]

    Phishing scams detection via tem- poral graph attention network in ethereum,

    L. Wang, M. Xu, and H. Cheng, “Phishing scams detection via tem- poral graph attention network in ethereum,” Information Processing & Management, vol. 60, no. 4, p. 103412, 2023

    2023

  28. [37]

    Siege: Self-supervised in- cremental deep graph learning for ethereum phishing scam detection,

    S. Li, R. Wang, H. Wu, S. Zhong, and F. Xu, “Siege: Self-supervised in- cremental deep graph learning for ethereum phishing scam detection,” in Proceedings of the 31st ACM International Conference on Multimedia , 2023, pp. 8881–8890

    2023

  29. [38]

    Tokenscout: Early detection of ethereum scam tokens via temporal graph learning,

    C. Wu, J. Chen, Z. Zhao, K. He, G. Xu, Y . Wu, H. Wang, H. Li, Y . Liu, and Y . Xiang, “Tokenscout: Early detection of ethereum scam tokens via temporal graph learning,” in Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security , 2024, pp. 956– 970

    2024

  30. [39]

    Tsgn: Transaction subgraph networks assisting phishing detection in ethereum,

    J. Wang, P. Chen, X. Xu, J. Wu, M. Shen, Q. Xuan, and X. Yang, “Tsgn: Transaction subgraph networks assisting phishing detection in ethereum,” IEEE Transactions on Dependable and Secure Computing , 2025

    2025

  31. [40]

    Grabphisher: Phishing scams detection in ethereum via temporally evolving gnns,

    J. Zhang, H. Sui, X. Sun, C. Ge, L. Zhou, and W. Susilo, “Grabphisher: Phishing scams detection in ethereum via temporally evolving gnns,” IEEE Transactions on Services Computing , 2024

    2024

  32. [41]

    Dynamic feature fusion: Combining global graph structures and local semantics for blockchain phishing detection,

    Z. Sheng, L. Song, and Y . Wang, “Dynamic feature fusion: Combining global graph structures and local semantics for blockchain phishing detection,” IEEE Transactions on Network and Service Management , pp. 1–1, 2025

    2025

  33. [42]

    The graph neural network model,

    F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfar- dini, “The graph neural network model,” IEEE transactions on neural networks, vol. 20, no. 1, pp. 61–80, 2008

    2008

  34. [43]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Ben- gio, “Graph attention networks,” arXiv preprint arXiv:1710.10903, 2017

    2017 arXiv

  35. [44]

    K-bert: Enabling language representation with knowledge graph,

    W. Liu, P. Zhou, Z. Zhao, Z. Wang, Q. Ju, H. Deng, and P. Wang, “K-bert: Enabling language representation with knowledge graph,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 34, no. 03, 2020, pp. 2901–2908

    2020

  36. [45]

    Translating embeddings for modeling multi-relational data,

    A. Bordes, N. Usunier, A. Garcia-Duran, J. Weston, and O. Yakhnenko, “Translating embeddings for modeling multi-relational data,” Advances in neural information processing systems , vol. 26, 2013

    2013

  37. [46]

    XBLOCK Blockchain Datasets: InPlusLab ethereum phishing detection datasets,

    L. Chen, J. Peng, Y . Liu, J. Li, F. Xie, and Z. Zheng, “XBLOCK Blockchain Datasets: InPlusLab ethereum phishing detection datasets,” http://xblock.pro/ethereum/, 2019

    2019

  38. [47]

    Xblock-eth: Extracting and exploring blockchain data from ethereum,

    P. Zheng, Z. Zheng, J. Wu, and H.-N. Dai, “Xblock-eth: Extracting and exploring blockchain data from ethereum,” IEEE Open Journal of the Computer Society, vol. 1, pp. 95–106, 2020

    2020

  39. [48]

    Etherscan labels: Full label data dump of top evm chains in json/csv,

    A. K. V . Brian, Command M, “Etherscan labels: Full label data dump of top evm chains in json/csv,” https://github.com/brianleect/ etherscan-labels, 2023, accessed: 2025-01-31

    2023

  40. [49]

    role2vec: Role-based network embeddings,

    N. K. Ahmed, R. A. Rossi, J. B. Lee, T. L. Willke, R. Zhou, X. Kong, and H. Eldardiry, “role2vec: Role-based network embeddings,” Proc. DLG KDD, pp. 1–7, 2019

    2019

  41. [50]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” arXiv preprint arXiv:1609.02907 , 2016

    2016 arXiv

  42. [51]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” Advances in neural information processing systems, vol. 30, 2017

    2017

  43. [52]

    Hierarchical graph representation learning with differentiable pooling,

    Z. Ying, J. You, C. Morris, X. Ren, W. Hamilton, and J. Leskovec, “Hierarchical graph representation learning with differentiable pooling,” Advances in neural information processing systems , vol. 31, 2018

    2018

  44. [53]

    Unsupervised Univer- sal Self-Attention Network for Graph Classification,

    D. Q. Nguyen, T. D. Nguyen, and D. Phung, “Unsupervised Univer- sal Self-Attention Network for Graph Classification,” arXiv preprint arXiv:1909.11855, 2019

    1909 arXiv

  45. [54]

    graph2vec: Learning distributed representations of graphs,

    A. Narayanan, M. Chandramohan, R. Venkatesan, L. Chen, Y . Liu, and S. Jaiswal, “graph2vec: Learning distributed representations of graphs,” arXiv preprint arXiv:1707.05005 , 2017

    2017 arXiv

  46. [55]

    Grabphisher: Phishing scams detection in ethereum via temporally evolving gnns,

    J. Zhang, H. Sui, X. Sun, C. Ge, L. Zhou, and W. Susilo, “Grabphisher: Phishing scams detection in ethereum via temporally evolving gnns,” IEEE Transactions on Services Computing , pp. 1–15, 2024

    2024

  47. [56]

    Variational graph auto-encoders,

    T. N. Kipf and M. Welling, “Variational graph auto-encoders,” arXiv preprint arXiv:1611.07308, 2016

    2016 arXiv

  48. [57]

    Graph attention auto-encoders,

    A. Salehi and H. Davulcu, “Graph attention auto-encoders,” arXiv preprint arXiv:1905.10715, 2019

    1905 arXiv

  49. [58]

    Roberta: A robustly optimized bert pretraining approach,

    Y . Liu, “Roberta: A robustly optimized bert pretraining approach,” arXiv preprint arXiv:1907.11692, vol. 364, 2019

    1907 arXiv

  50. [59]

    Quaternion knowledge graph em- beddings,

    S. Zhang, Y . Tay, L. Yao, and Q. Liu, “Quaternion knowledge graph em- beddings,” Advances in neural information processing systems , vol. 32, 2019

    2019

  51. [60]

    Com- plex embeddings for simple link prediction,

    T. Trouillon, J. Welbl, S. Riedel, É. Gaussier, and G. Bouchard, “Com- plex embeddings for simple link prediction,” in International conference on machine learning . PMLR, 2016, pp. 2071–2080

    2016

  52. [61]

    Rotate: Knowledge graph embedding by relational rotation in complex space,

    Z. Sun, Z.-H. Deng, J.-Y . Nie, and J. Tang, “Rotate: Knowledge graph embedding by relational rotation in complex space,” arXiv preprint arXiv:1902.10197, 2019

    1902 arXiv

  53. [62]

    Embedding entities and relations for learning and inference in knowledge bases,

    B. Yang, W.-t. Yih, X. He, J. Gao, and L. Deng, “Embedding entities and relations for learning and inference in knowledge bases,” arXiv preprint arXiv:1412.6575, 2014

    2014 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.