# From Viral to Void: Multi-Dimensional Behavioral and Contractual Analysis for Rug Pull Identification

Crypto.world

> Rug pull scams in meme coins let a project team attract liquidity and then abruptly drain it, leaving investors with worthless tokens. Existing detection efforts tend to use a single feature…

**arXiv**: 2608.01609

**Published**: 2026-08-03

**Authors**: Jinyin Song, Hongping Wang, Xiaoqi Li

## AI summary

## Overview

**Research area:** Blockchain security and DeFi fraud detection — specifically identifying "rug pull" scams in Ethereum-ecosystem meme coins through combined smart-contract analysis, on-chain transaction behavior, and machine learning.

**Technical level:** Intermediate. The paper is readable with a basic grasp of smart contracts, liquidity pools, and standard ML classification concepts; the mathematical content (constant-product pricing, Gini coefficient, Herfindahl-Hirschman Index) is presented at an introductory level.

**Scope (one sentence):** The paper defines rug pull mechanics, builds a multi-dimensional feature system from contract code and on-chain behavior, trains an MLP classifier with imbalance-handling techniques, validates the framework against a documented BSC case ("Second Uncle Coin"), and packages the result as a Flask-based web detection system.

## What This Paper Is About

Rug pull scams in meme coins let a project team attract liquidity and then abruptly drain it, leaving investors with worthless tokens. Existing detection efforts tend to use a single feature dimension, handle class imbalance poorly, and stay at the offline-experiment stage without usable tooling. This paper aims to combine contract-level, transaction-level, liquidity-level, and social-media features into one interpretable detection pipeline, and to deliver it as a deployable early-warning tool rather than a purely experimental model.

## Key Contributions

1. **A multi-dimensional feature system for rug pull detection.** The authors construct features across four dimensions: malicious smart contract design (backdoors, privilege abuse, proxy traps, hidden minting), on-chain transaction anomalies, liquidity manipulation, and social media disclosures — with quantitative metrics and explicit anomaly thresholds for each behavior category (Table 1).
2. **A formalized taxonomy and mechanism analysis.** The paper separates rug pulls into liquidity drain, abuse of privileges, and token backdoor attack types, plus "hard pull" versus "soft pull" by duration and scale, and derives a deploy-time risk scoring function *R_deploy* from Boolean contract indicators and locked-supply ratio.
3. **An MLP-based risk detector with combined imbalance handling.** The model pairs SMOTE oversampling with Focal Loss, uses dynamic threshold search based on optimal F1 on the validation set instead of a fixed 0.5 cutoff, and adds gradient pruning and early stopping for training stability.
4. **A deployable visualization system.** A Flask web application provides batch risk assessment, high-risk ranking display, and result export, moving the work beyond offline experimentation.

## Main Findings

- **Test-set performance:** The model reaches an accuracy of 0.927, an F1 score of 0.787, and an AUC-ROC of 0.952 on the test set, which the authors report as outperforming traditional methods (the specific baseline methods are not named in the provided content).
- **Case reconstruction of "Second Uncle Coin" (BOBU):** Deployed on Binance Smart Chain on July 28, 2022, the token had an initial total supply of 1 billion. The deployer's address received approximately 40%, another 30% was split between two additional controlled addresses, and the remaining 30% stayed in the contract for liquidity. The entire sequence — deployment, hype, price manipulation, liquidity drain — was completed in two days, netting approximately $1.3 million.
- **Contract-level red flags in the case:** The contract contained `setTaxRate` with no time lock or upper limit (allowing tax from 0% to 100%), an `addToBlacklist`/`removeFromBlacklist` mechanism, and a reserved upgrade function. With weights w1 = w2 = w3 = 0.3 and w4 = 0.1, and lockedRatio = 0.3, the derived score is R_deploy = 0.97, versus a typical project threshold usually below 0.3.
- **Liquidity provision arithmetic:** 300 million BOBU were paired with approximately 500 BNB (about $120,000 at the time) on PancakeSwap V2, one hour after deployment and with no third-party locking platform. This gives k = 1.5 × 10^11 and an initial price P0 ≈ 1.67 × 10^-6 BNB/BOBU.
- **Quantitative detection thresholds:** Initial liquidity more than 10 times the median of similar projects; lock-up duration under 30 days; withdrawal ratio above 90%; volume-padding index W > 10; cross-address transaction-amount correlation ρ > 0.9; more than 10 ETH sent to a mixer within 1 hour of withdrawal; HHI > 0.25; Gini coefficient G > 0.8 for rug pulls versus below 0.5 for normal projects.
- **Holding-concentration contrast:** A healthy decentralized token's top ten addresses typically hold no more than 30% of total supply, while in a rug pull project team-controlled addresses may collectively hold more than 60%, with the proportions shifting significantly on the eve of an attack.
- **Class imbalance as a structural problem:** Rug pull projects often account for less than 5% of actual data, which biases a naively trained classifier toward the "safe" class and produces very low recall — the motivation for the SMOTE plus Focal Loss combination.
- **Mechanism explanation for the price collapse:** Under the constant-product AMM model, when the team withdraws Δx close to x and Δy close to y, the new price P' either approaches 0 or infinity; in practice remaining ETH approaches 0 and the token price plummets to near 0, so even small sells return x_out ≈ 0 and transactions fail.

## Methodology in Plain English

The authors start with definitions. They describe how a rug pull works mechanically — a team seeds a liquidity pool on an automated market maker such as Uniswap or PancakeSwap, builds hype, then removes the liquidity or abuses a contract privilege. From this they derive observable signals.

They then select a documented real-world case, "Second Uncle Coin," and reconstruct its lifecycle stage by stage: token deployment, liquidity provision, fake trading, and fund withdrawal. Each stage is translated into a measurable quantity — for example, the fraction of supply held by team-controlled addresses at deployment, the ratio of withdrawn to deposited liquidity, the regularity of transaction intervals, and the Gini coefficient of holder balances. These become the feature set.

The detection model is a Multi-Layer Perceptron. Because fraudulent projects are a small minority of the data (under 5%), the training pipeline generates synthetic minority examples with SMOTE and reshapes the loss with Focal Loss so the model is penalized more for missing actual rug pulls. Rather than accepting the default 0.5 probability cutoff, the authors search for the threshold that maximizes F1 on validation data, which trades off precision against recall deliberately. Gradient pruning and early stopping are used to keep training stable.

Finally, the trained model is wrapped in a Flask web application so a user can submit tokens for batch scoring, view a ranked list of high-risk projects, and export the results.

## Why This Matters

**Impact on research.** The work argues that rug pull detection needs genuinely multi-dimensional evidence — contract semantics, liquidity state, transaction graph patterns, and social signals — rather than transaction-network analysis alone. It also foregrounds two methodological points that the authors identify as gaps in prior work: imbalanced learning at the sub-5% positive rate, and dynamic threshold selection replacing the fixed 0.5 decision boundary. The explicit threshold table gives other researchers reproducible criteria to test against.

**Real-world applications:**

- **Retail investor screening:** a buyer can batch-check a new meme coin for hidden minting, blacklist functions, unlocked liquidity, and concentrated holdings before committing funds.
- **Exchange and DEX listing review:** platforms can rank newly deployed tokens by risk score and gate listings or flag pools for review.
- **Regulatory and forensic monitoring:** the fund-consolidation and mixer-interaction indicators (large transfers to mixing services shortly after withdrawal) support post-hoc investigation and case building.
- **Wallet and aggregator integration:** a risk score can be surfaced as a warning at the point of swap.

**Industry relevance.** Because token deployment on chains like Ethereum and BSC is nearly free and anonymous, and because post-theft funds move through cross-chain bridges and mixers such as Tornado Cash, detection has to be fast and automated. The paper targets this directly by producing a running web system rather than a static model, and it notes that the marketing tactics involved — forged audit reports, exaggerated team credentials, paid promotion — are cheap to produce and hard to verify manually.

## Future Directions

- **Extending coverage beyond a single chain:** the authors note that detection methods covering multiple public blockchains such as ETH, BSC, and SOL remain relatively rare; generalizing the feature system across chains is an open problem.
- **Operationalizing the social-media dimension:** social signals are part of the stated feature system, but the paper does not report how sentiment or disclosure features are quantified or how much they contribute to the model's performance.
- **Validating the thresholds at scale:** the anomaly cutoffs (W > 10, ρ > 0.9, G > 0.8, HHI > 0.25, and so on) are motivated by the single reconstructed case and general reasoning; testing them against a large labeled corpus would establish how well they generalize.
- **Improving recall and interpretability further:** with F1 at 0.787 against accuracy of 0.927, there is room to reduce false negatives on the minority class, and the paper identifies weak model interpretability as an unresolved issue in the field.

## Target Audience

Blockchain security researchers and graduate students working on fraud detection in DeFi; smart contract auditors who want a checklist of contract-level risk indicators; ML practitioners applying imbalanced classification to financial crime data; and builders of consumer-facing crypto safety tools who need a concrete feature set and deployment pattern. Investors with moderate technical literacy will find the case reconstruction and threshold table the most immediately useful parts.

*Note: the provided content is truncated and does not report the dataset size, the names of the baseline methods used for comparison, or the detailed experimental setup; those details would need to be read from the full paper.*

## Authors’ abstract

As the blockchain and decentralized finance (DeFi) ecosystems continue to expand and mature, rug pull scams involving meme coins are occurring with increasing frequency, posing a threat to the security of investors' assets and the healthy development of the industry. Rug Pull scams are characterized by extremely low deployment costs, covert execution, rapid fund transfers, and high detection difficulty. Traditional manual reviews or fixed rules struggle to meet real-time…

[The full abstract on arXiv](https://arxiv.org/abs/2608.01609)

[Read the original paper](https://arxiv.org/abs/2608.01609)

---

Canonical: https://crypto.world/paper/from-viral-to-void-multi-dimensional-behavioral-and-contractual-2608.01609
