Skip to main content

Table 1 Features used in related studies

From: Detecting malicious accounts in permissionless blockchains using temporal graph properties

# B/C Used features based on (See Abbreviation section) ML Algo used Dataset Hyperparameters Performance
AS iD oD Bal TF BB A CC IET
Pham and Lee (2016) B K-means 100\(\hbox {K}^a\) \(k \in [1,14]\) \(k_{opt}=7,8\)
Mahalanobis distance \(\times\) 0.0256\(^{MDE}\)
\(\nu\)-SVM \(\nu =0.005\) 0.1441\(^{MDE}\)
Pham and Lee (2017) B - Local outlier factor 6.3\(\hbox {M}^a\) \(k = 8\) 0.55\(^{MDE}\)
Monamo et al. (2016) B K-means 1\(\hbox {M}^a\) \(k \in [1,14]\) \(k_{opt}=8\)
Trimmed K-means \(k \in [1,15]\), \(\alpha =0.01\) \(k_{opt}=8\)
Bartoletti et al. (2018) B RIPPER\(\dagger\) \(\ddagger\)6432\(^a\) Cost \(\in [1,40]\) 0.996\(^{ac}\)
Bayes network \(\times\) 0.983\(^{ac}\)
Random Forest \(\times\) 0.996\(^{ac}\)
Chen et al. (2018b) E XGBoost \(\ddagger\)1382\(^{sc}\) \(\times\) 0.94\(^p\), 0.81\(^r\)
Ostapowicz and Zbikowski (2019) E Random Forest 350\(\hbox {K}^a\) RFPARAM 0.85\(^{r}\), 0.05\(^{p}\)
SVM \(Cost=1\), \(\gamma =0.077\) 0.87\(^{r}\), 0.02\(^{p}\)
XGBoost XGBPARAM 0.8\(^{r}\), 0.07\(^{p}\)
Singh (2019) E Decision Tree 300\(^a\) \(\times\) 0.93\(^{ac}\)
SVM \(\times\) 0.83\(^{ac}\)
KNN \(k=5\) 0.91\(^{ac}\)
MLP \(\times\) 0.86\(^{ac}\)
NaiveBayes \(\times\) 0.89\(^{ac}\)
Random Forest \(\times\) 0.99\(^{ac}\)
Kumar et al. (2020) E Decision Tree 9375\(^a\) \(\times\) 0.92\(^{ac}\)
KNN \(\times\) 0.92\(^{ac}\)
XGBoost \(\times\) 0.96\(^{ac}\)
Random Forest \(\times\) 0.95\(^{ac}\)
Zola et al. (2019) B Adaboost 1000\(\hbox {M}^a\) \(Estimators=50\), \(rate=1\) \(>0.2^{r}\)
Random Forest \(Estimators=10\) \(>0.85^{r}\)
Gradient boosting \(estimators=100\), \(rate=0.1\) \(>0.93^{r}\)
\(Depth=3\)
  1. \({}^{B/C}\) Blockchain, \({}^B\) Bitcoin, \({}^E\) Ethereum, \({}^a\) accounts, \({}^{sc}\) Smart Contracts, \({}^{MDE}\) Dual Evaluation Metric, \({}^{ac}\) accuracy, \({}^p\) Precision, \({}^r\) Recall, \({}^{\dagger }\) it is a propositional rule learner that relies on a sequential covering logic, \({}^{\ddagger }\) Ponzi scheme data, \({}^{RFPARM}\) features = 3, leaf samples = 10, threshold probability = 0.99, \({}^{XGBPARAM}\) depth = 3, child weight = 8, subsample = 1, probability = 0.99, \({}^{\times }\) not provided