Detecting Fraudulent Accounts on Blockchain: A Supervised Approach

Kamil \.Zbikowski; Michal Ostapowicz

REVIEW 4 major objections 7 minor 12 references

Detecting Fraudulent Accounts on Blockchain: A Supervised Approach

T0 review · 4 major / 7 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A Random Forest trained on 13 Ethereum transaction aggregates flags known fraud accounts with a 0.02% false-positive rate while catching 23.67% of them, and the authors propose it as an automated anti-fraud rule for wallets and exchanges.

desk verdict Honest Ethereum fraud-detection benchmark, but the headline deployment claim rests on a look-ahead bias the authors acknowledge and never resolve. read the letter →

arxiv 1908.07886 v1 pith:VYFKVJ6K submitted 2019-08-21 cs.CR

Michal Ostapowicz , Kamil Żbikowski This is my paper · ORCID

classification cs.CR

keywords ethereumblockchainfrauddetectionsupervisedlearningrandomforestxgboostsupportvectormachinelook-aheadbias

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 argues that supervised machine learning can identify fraudulent Ethereum accounts from a small set of transaction aggregates, and that one Random Forest configuration is precise enough to serve as an automated anti-fraud rule in real wallets or exchanges. The authors train Random Forests, Support Vector Machines, and XGBoost on 2,200 wallets publicly tagged as 'Hack/Phishing' plus 349,999 randomly selected untagged wallets. Their best precision-oriented model flags only 0.02% of honest accounts while still catching 23.67% of known frauds, and their best recall model finds 84.92% of frauds at the cost of a 9.69% false-positive rate. They also run a sensitivity analysis to show how performance degrades when the most important transaction features are removed, which is their main response to the risk that post-fraud activity contaminates the training data.

What carries the argument

The central object is a 13-feature profile built from on-chain transaction data: counts of incoming and outgoing transactions, unique counterparts, average and total values, average time gaps between incoming and outgoing transactions, average gas price, average gas limit, and active lifetime in days. Each Ethereum address is reduced to this vector, and the three classifiers separate fraudulent from non-fraudulent profiles on that basis. The load-bearing evaluation choice is the false-positive rate instead of precision, because the sample is artificially balanced and precision would therefore overstate real-world usefulness; the false-positive rate is independent of the inflated fraud share in the training set.

What would settle it

Take the accounts tagged as fraudulent on the blockchain explorer, find the date each tag was added from the platform's history, and recompute the 13 features using only transactions before that date; if the Random Forest's 0.02% false-positive rate and 23.67% recall are not reproduced on this look-ahead-free dataset, the reported performance depends on post-exposure transactions. A second check would be to manually audit a random sample of the 349,999 'honest' wallets: if a non-negligible share are actually phishing or hack addresses, the false-positive rate is understated.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central result is that the Random Forest with three random features per split, a minimum node size of ten, and a 0.5 probability threshold (Configuration 3) achieves 99.98% specificity and a 0.02% false-positive rate on a held-out validation set while detecting 23.67% of fraud-labeled accounts, a combination the authors propose as a good candidate for an automated anti-fraud system. A high-recall companion (Configuration 19) detects 84.92% of frauds but flags 9.69% of all accounts, which the authors judge too noisy for real-world deployment. Across all three algorithms the probability threshold controls the trade-off between recall and false positives, and the single most important feature is the average time between incoming transactions.

Load-bearing premise

The load-bearing premise is that the blockchain explorer's user-reported 'Hack/Phishing' tags are accurate ground truth and that the 349,999 randomly sampled unmarked wallets contain essentially no fraud; the paper also assumes its 13 aggregates are computed only from data available before each account was publicly exposed, an assumption it admits it cannot verify.

Editorial extensions

If this is right

If the 0.02% false-positive rate holds in production, a wallet or exchange could run this Random Forest as a pre-screen: roughly one in five thousand honest accounts would be wrongly flagged while about one in four fraud accounts would be caught automatically.
The high-recall configuration, despite its 84.92% recall, is not suitable for fully automated deployment because nearly 10% of all accounts would be alerted, overwhelming any manual review process.
Because the method relies only on publicly visible transaction aggregates, the same features can be computed for other blockchains such as Bitcoin, as the paper states the approach is easily transferable.
The sensitivity analysis shows that removing the most important variables degrades recall much more than specificity: the low-FPR configuration stays at 0.02% false positives but catches only 7.66% of frauds when the eight top features are dropped, meaning the model's coverage depends heavily on those features.

Reading between the lines

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

A testable extension the paper does not explore: compute the 13 features from only the first few days of an account's life and run the low-FPR model as a sandbox filter for newly created accounts, so that fraud is blocked before large funds arrive; this would need to be validated against a time-split dataset.
The prominence of average time between incoming transactions suggests a behavioural signature: fraud accounts receive funds in short bursts soon after creation and then go quiet. A simple rule on transaction-timing variance could be tested as a lightweight complement to the classifier.
Because the ground-truth tags are user reports, the true rate of missed fraud is unknown. A natural follow-up is to apply the trained model to accounts that were later added to the tag list and measure how many had already been flagged before the public report appeared.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper is an honest supervised-learning benchmark for flagging fraud-labeled Ethereum accounts. The headline result is Random Forest Conf. 3: 0.02% false positive rate and 23.67% recall on a held-out validation set. That is a useful number if the features are pre-exposure. The problem: they almost certainly are not, and the authors know it.

What's new: three off-the-shelf classifiers on 13 transaction aggregates for 350k+ accounts, with a sensitivity analysis that drops high-importance variables. Using false positive rate instead of precision is the right call given the inflated class ratio in the validation set. Credit is fairly given to Pham and Lee for the feature set.

Where it is soft, in proportion: the load-bearing flaw is look-ahead bias. Section 3.2 admits the data do not reveal when a wallet was flagged, so aggregates may include post-exposure transactions. For a wallet publicly tagged as fraudulent, post-exposure incoming transfers are exactly the traffic a scam generates, so features like total incoming value can leak the label. The sensitivity analysis removes the top n features, but that shows graceful degradation, not the absence of a leak; remaining features can still encode post-exposure behavior. The authors relegate 'estimating exact moments' to future work. So the conclusion that Conf. 3 is 'a good candidate for an automated anti-fraud system' is unsupported by the evidence as presented. A temporal holdout that computes features only up to exposure time would settle it, and it is not here.

Second, labels are user-reported Hack/Phishing tags, and the random wallets are assumed clean. That undercounts fraudulent accounts and makes the 0.02% FPR a lower bound under optimistic assumptions.

The math and modeling are otherwise standard and fine. The paper is clearly written and the limitation discussion is candid.

Who is this for? Practitioners building blockchain screening tools will get a working pipeline and feature comparison, if they treat the deployment claim with caution. Researchers studying temporal leakage in ledger data may also find it a useful case study.

My recommendation: yes, send to peer review. The contribution is a reasonable benchmark, the flaw is real but addressable, and expert feedback could push the authors either to fix the temporal contamination or restate the conclusion.

Referee Report

4 major / 7 minor

Summary. The paper presents a supervised machine-learning pipeline for detecting fraudulent accounts on the Ethereum blockchain. The authors collect 2,200 Etherscan user-tagged "Hack/Phishing" wallets and 349,999 randomly selected wallets as non-fraudulent, derive 13 transaction-level aggregate features, and train Random Forest, SVM, and XGBoost classifiers using grid search with 10-fold cross-validation. They select configurations that trade off recall and false-positive rate, report validation-set results, and perform a sensitivity analysis in which the most important features are removed. The central claim is that Random Forest configuration 3, with a 0.02% false-positive rate and 23.67% recall on the validation set, is "a good candidate for an automated anti-fraud system" (Section 5).

Significance. The paper is a solidly executed proof-of-concept with several strengths: the dataset is large and publicly observable, the comparison covers three standard algorithms, the use of false-positive rate rather than precision is a reasonable response to the artificial class distribution, and the sensitivity analysis directly addresses the acknowledged look-ahead risk. However, the application-oriented conclusion is not currently supported: the unresolved look-ahead contamination in the transaction aggregates, the noisy user-reported ground truth, and the absence of uncertainty quantification all affect the headline numbers. If the authors can reframe the contribution as an offline detection benchmark or obtain pre-exposure data, the work would be a useful baseline for blockchain fraud detection. As written, the paper's contribution is a promising method rather than a validated deployment-ready system.

major comments (4)

[Section 3.2, Section 4.4, Tables 2-4, Figure 2] Section 3.2 states that the data do not allow determining when an account was marked as fraudulent and that some aggregates may include transactions made after the account's public exposure. For a fraud account, that post-exposure period is precisely when victims are likely to send funds, so features such as VIT, AVIT, and ATIT (Table 1), which Figure 2 identifies as the most important variables, can be contaminated with information unavailable at the time a screening decision would be made. The sensitivity analysis in Section 4.4 removes the n most important variables, but this does not eliminate the leak: the remaining variables can still encode post-exposure behavior, and the experiment only shows how the fitted model degrades, not how it would perform on features computed from pre-exposure data. The held-out FPR and recall estimates in Tables 2-4 therefore do not establish the Section 5 claim that Conf. 3 is a good candidate for an automated anti-fraud system. The authors should reconstruct features using the earliest available timestamp of the fraud label as a proxy for exposure time, or explicitly restrict the claim to post-hoc identification of known-fraud accounts.
[Section 3.1] The ground-truth labels are taken from Etherscan's user-reported "Hack/Phishing" tags, and the negative class consists of 349,999 randomly selected wallets that are simply "not marked as suspicious." User-reported tags can contain false positives, unmarked wallets can contain unreported frauds, and no manual verification is described. Because every reported recall, precision, and FPR value is computed against this labeling, none of the headline numbers can be interpreted cleanly as the probability of detecting a true fraud or of avoiding a false alarm. At minimum, the authors should report a manual audit of a random sample of both classes, or rerun the evaluation on a curated set of confirmed fraud accounts and an explicitly vetted negative set, and discuss how label noise would change the operating point.
[Section 3.2, Section 4.3] The experiments search a large number of hyperparameter configurations (20 RF, 20 SVM, and 240 XGBoost configurations per Section 4.3) and then report the best configuration for each metric on the same validation set. The paper gives no confidence intervals, bootstrap estimates, or a separate test set used only once, so the reported point estimates, especially the 0.02% FPR of Conf. 3, are likely optimistic to an unknown degree. The authors should provide repeated split or bootstrap intervals for the main configurations, or at minimum state how stable the ranking is across random seeds and CV folds.
[Section 5, Table 2] The deployment argument in Section 5 needs an explicit false-discovery-rate calculation under realistic Ethereum fraud prevalence. Because the validation set is not representative of the deployment population, the 85.71% precision of Conf. 3 in Table 2 is not directly usable. The paper's statement that "one in five thousands accounts" would be marked as fraudulent refers only to the false-positive rate among non-fraud accounts; with a realistic baseline fraud prevalence of, for example, 0.1%, the expected proportion of flagged accounts is roughly 0.044% (about 1 in 2,300), of which only about half are true frauds, and at lower prevalence the precision drops correspondingly. Without this analysis, the "good candidate for an automated anti-fraud system" claim is not quantitatively supported.

minor comments (7)

[Section 4.1] In the paragraph introducing cut-off probabilities, "predicted as a non-fraud" should read "predicted as a fraud"; the surrounding text and Table 2 show that higher thresholds increase recall, which corresponds to flagging more accounts as fraudulent.
[Tables 2, 5, and 6] The column header "Cross-validation results [%]" conflicts with the table captions "Validation results"; these tables report the held-out validation metrics, not cross-validation results.
[Section 4.4] The sentence about XGBoost's "minor change... 2000 iterations regardless" is vague; specify whether early stopping is disabled in the sensitivity analysis and whether that affects overfitting.
[Abstract and Section 5] The abstract says recall and precision values "allow for the designed system to be applicable as an anti-fraud rule," but for the low-FPR configuration the recall is 23.67%; the authors should state how this level was deemed sufficient for the intended application.
[Section 3.1] Clarify whether the 349,999 random wallets were subjected to the same activity filter as the positive wallets; if not, the classifier may be distinguishing active from inactive accounts rather than fraud from non-fraud.
[General] No code or dataset availability is mentioned; given that the data are publicly retrievable from Etherscan, including the preprocessing scripts would materially improve reproducibility.
[General] The manuscript contains several typos (for example, "Etherum" in Section 1 and "fraudulent transaction" in Section 4); a careful proofread is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the supervised learning evaluation is self-contained and out-of-sample; the acknowledged look-ahead bias is a data-contamination concern, not circular reasoning.

full rationale

The paper's central claim is an empirical comparison of Random Forest, SVM, and XGBoost classifiers on a held-out validation set. Features are 13 transaction aggregates computed from public Ethereum blockchain data, and labels come from Etherscan's user-reported 'Hack/Phishing' tags plus random unmarked wallets. The reported results (e.g., Random Forest Conf. 3 with 0.02% FPR and 23.67% recall) are measured predictions against the held-out labels, not quantities derived from the model inputs by construction. Hyperparameters are selected by grid search with cross-validation, but the final validation metrics are genuine out-of-sample estimates. The paper contains no self-citations that carry a load-bearing argument, no uniqueness theorem imported from the authors' prior work, and no fitted parameter renamed as a prediction. The authors explicitly and repeatedly acknowledge that the exact time of fraud labeling is unknown and that transaction aggregates may include post-exposure transactions; they call this a possible 'look-ahead bias' and treat it as a limitation and future work. That is a real threat to external validity and to the practical claim about an automated anti-fraud system, but it is not a circularity of the derivation: the evaluation pipeline still compares model outputs with independent labels on data not used for training. The statement that this result 'can be perceived as a good candidate for an automated anti-fraud system' is a forward-looking interpretation, not a mathematically forced consequence of the paper's definitions. Therefore no step in the claimed derivation reduces to its own inputs.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the Etherscan labels being accurate, the random sample representing true non-fraud, the sufficiency of the 13 engineered features, and the chosen hyperparameters and classification thresholds. The paper explicitly acknowledges the look-ahead bias and the unreliability of precision due to class imbalance, but does not provide external verification of labels or a temporal validation, so these assumptions remain load-bearing.

free parameters (6)

Classification probability threshold (RF Conf. 3) = 0.5
Chosen by grid search; controls the recall versus false-positive trade-off.
mtry (RF Conf. 3) = 3
Grid-searched; number of random features per split.
min.node.size (RF Conf. 3) = 10
Grid-searched; minimum leaf size.
Classification probability threshold (RF Conf. 19) = 0.99
Grid-searched; high-recall setting.
SVM cost and gamma = cost 1-50, gamma 0.077-2 (grid)
Grid-searched; SVM results reported across the grid.
XGBoost max.depth, colsample, min.child.weight = max.depth 3-9, colsample 0.25-1.0, min.child.weight 1-8 (grid)
Grid-searched over 240 combinations; the paper reports 20 selected configurations.

assumptions (4)

domain assumption Etherscan 'Hack/Phishing' user reports are an accurate ground truth for fraudulent accounts.
Section 3.1: The target labels come from user-reported tags on Etherscan; the paper does not verify them against judicial or exchange records. Noisy or incomplete labels would bias all reported metrics.
domain assumption Randomly selected wallets not tagged as suspicious are non-fraudulent.
Section 3.1: 349,999 random wallets are treated as the negative class; some may be fraudulent but not yet tagged, which would inflate false positives in the evaluation.
domain assumption The 13 transaction aggregate features are sufficient to distinguish fraudulent from non-fraudulent accounts.
Section 3.1 Table 1: The feature set is inspired by prior work but is a modeling choice; the paper provides no argument that these features capture all relevant fraud signals.
standard math Standard ML algorithm correctness (SVM, Random Forest, XGBoost).
The paper relies on well-known algorithms as cited; we accept their correctness mutatis mutandis.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Detecting Fraudulent Accounts on Blockchain: A Supervised Approach." pith.science (2026). https://pith.science/paper/VYFKVJ6K

@misc{pith2026190807886,
  author       = {Pith},
  title        = {Pith review of: Detecting Fraudulent Accounts on Blockchain: A Supervised Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VYFKVJ6K}},
  note         = {Machine review of arXiv:1908.07886}
}

read the original abstract

Applications of blockchain technologies got a lot of attention in recent years. They exceed beyond exchanging value and being a substitute for fiat money and traditional banking system. Nevertheless, being able to exchange value on a blockchain is at the core of the entire system and has to be reliable. Blockchains have built-in mechanisms that guarantee whole system's consistency and reliability. However, malicious actors can still try to steal money by applying well known techniques like malware software or fake emails. In this paper we apply supervised learning techniques to detect fraudulent accounts on Ethereum blockchain. We compare capabilities of Random Forests, Support Vector Machines and XGBoost classifiers to identify such accounts basing on a dataset of more than 300 thousands accounts. Results show that we are able to achieve recall and precision values allowing for the designed system to be applicable as an anti-fraud rule for digital wallets or currency exchanges. We also present sensitivity analysis to show how presented models depend on particular feature and how lack of some of them will affect the overall system performance.

Figures

Figures reproduced from arXiv: 1908.07886 by the authors.

**Figure 1.** System and data architecture for the conducted experiment 3.3 Prediction models The Support Vector Machine (SVM) classifier is a binary classifier algorithm that looks for an optimal hyperplane as a decision function in a high-dimensional [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

**Figure 2.** Variable importance for: a) XGBoost Conf. 1, b) XGBoost Conf. 16, c) Random Forest Conf. 3 [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 7 canonical work pages

[1]

Journal of Network and Computer Applications 68, 90–113 (2016)

Abdallah, A., Maarof, M.A., Zainal, A.: Fraud detection system: A survey. Journal of Network and Computer Applications 68, 90–113 (2016)

2016
[2]

International Journal of Computer Applications 156(10) (2016)

Bhardwaj, A., Gupta, R.: Financial frauds: Data mining based detection–a com- prehensive survey. International Journal of Computer Applications 156(10) (2016)

work page 2016
[3]

In: Proceedings of the ﬁfth annual workshop on Computational learning theory

Boser, B.E., Guyon, I.M., Vapnik, V.N.: A training algorithm for optimal margin classiﬁers. In: Proceedings of the ﬁfth annual workshop on Computational learning theory. pp. 144–152. ACM (1992)

1992
[4]

Machine learning 45(1), 5–32 (2001)

Breiman, L.: Random forests. Machine learning 45(1), 5–32 (2001)

2001
[5]

white paper (2014)

Buterin, V., et al.: A next-generation smart contract and decentralized application platform. white paper (2014)

work page 2014
[6]

Decision Support Systems 95, 91 – 101 (2017)

Carneiro, N., Figueira, G., Costa, M.: A data mining based system for credit-card fraud detection in e-tail. Decision Support Systems 95, 91 – 101 (2017). https://doi.org/https://doi.org/10.1016/j.dss.2017.01.002, http:// www.sciencedirect.com/science/article/pii/S0167923617300027

work page doi:10.1016/j.dss.2017.01.002 2017
[7]

CoRR abs/1603.02754 (2016), http://arxiv.org/abs/1603.02754 14 Micha l Ostapowicz and Kamil ˙Zbikowski

Chen, T., Guestrin, C.: Xgboost: A scalable tree boosting system. CoRR abs/1603.02754 (2016), http://arxiv.org/abs/1603.02754 14 Micha l Ostapowicz and Kamil ˙Zbikowski

arXiv 2016
[8]

In: IEEE International Conference on Networking, Sensing and Con- trol, 2004

Kou, Y., Lu, C.T., Sirwongwattana, S., Huang, Y.P.: Survey of fraud detection techniques. In: IEEE International Conference on Networking, Sensing and Con- trol, 2004. vol. 2, pp. 749–754. IEEE (2004)

work page 2004

Show all 12 references

[9]

CoRR abs/1611.03941 (2016), http://arxiv.org/abs/1611

Pham, T., Lee, S.: Anomaly detection in bitcoin network using unsupervised learning methods. CoRR abs/1611.03941 (2016), http://arxiv.org/abs/1611. 03941

2016 arXiv
[10]

Expert Systems with Applications 35(4), 1721 – 1732 (2008)

Quah, J.T., Sriganesh, M.: Real-time credit card fraud detection using com- putational intelligence. Expert Systems with Applications 35(4), 1721 – 1732 (2008). https://doi.org/https://doi.org/10.1016/j.eswa.2007.08.093, http://www. sciencedirect.com/science/article/pii/S09574...

2008 doi
[11]

Ethereum project yellow paper 151, 1–32 (2014)

Wood, G., et al.: Ethereum: A secure decentralised generalised transaction ledger. Ethereum project yellow paper 151, 1–32 (2014)

2014
[12]

W¨ orner, D., Von Bomhard, T., Schreier, Y.P., Bilgeri, D.: The bitcoin ecosystem: disruption beyond ﬁnancial services? (2016)

2016

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Journal of Network and Computer Applications 68, 90–113 (2016)

Abdallah, A., Maarof, M.A., Zainal, A.: Fraud detection system: A survey. Journal of Network and Computer Applications 68, 90–113 (2016)

2016

[2] [2]

International Journal of Computer Applications 156(10) (2016)

Bhardwaj, A., Gupta, R.: Financial frauds: Data mining based detection–a com- prehensive survey. International Journal of Computer Applications 156(10) (2016)

work page 2016

[3] [3]

In: Proceedings of the ﬁfth annual workshop on Computational learning theory

Boser, B.E., Guyon, I.M., Vapnik, V.N.: A training algorithm for optimal margin classiﬁers. In: Proceedings of the ﬁfth annual workshop on Computational learning theory. pp. 144–152. ACM (1992)

1992

[4] [4]

Machine learning 45(1), 5–32 (2001)

Breiman, L.: Random forests. Machine learning 45(1), 5–32 (2001)

2001

[5] [5]

white paper (2014)

Buterin, V., et al.: A next-generation smart contract and decentralized application platform. white paper (2014)

work page 2014

[6] [6]

Decision Support Systems 95, 91 – 101 (2017)

Carneiro, N., Figueira, G., Costa, M.: A data mining based system for credit-card fraud detection in e-tail. Decision Support Systems 95, 91 – 101 (2017). https://doi.org/https://doi.org/10.1016/j.dss.2017.01.002, http:// www.sciencedirect.com/science/article/pii/S0167923617300027

work page doi:10.1016/j.dss.2017.01.002 2017

[7] [7]

CoRR abs/1603.02754 (2016), http://arxiv.org/abs/1603.02754 14 Micha l Ostapowicz and Kamil ˙Zbikowski

Chen, T., Guestrin, C.: Xgboost: A scalable tree boosting system. CoRR abs/1603.02754 (2016), http://arxiv.org/abs/1603.02754 14 Micha l Ostapowicz and Kamil ˙Zbikowski

arXiv 2016

[8] [8]

In: IEEE International Conference on Networking, Sensing and Con- trol, 2004

Kou, Y., Lu, C.T., Sirwongwattana, S., Huang, Y.P.: Survey of fraud detection techniques. In: IEEE International Conference on Networking, Sensing and Con- trol, 2004. vol. 2, pp. 749–754. IEEE (2004)

work page 2004