LROO Rug Pull Detector: A Leakage-Resistant Framework Based on On-Chain and OSINT Signals

TL;DR

This paper introduces LROO Rug Pull Detector, a leakage-resistant framework that combines on-chain metrics and OSINT signals for early, reliable detection of rug pulls in smart contract ecosystems, improving over existing reactive methods.

Contribution

It presents a novel leakage-aware, multimodal detection framework using temporally aligned on-chain and OSINT data, with a curated dataset and causal modeling approach for early fraud detection.

Findings

Achieves high discriminative performance in detecting rug pulls.

Maintains low false-negative rates in early detection.

Demonstrates improved calibration over classical baselines.

Abstract

Smart contract-based ecosystems enable decentralized applications without trusted intermediaries, but their immutability and permissionless design also facilitate large-scale fraud. One of the most prevalent attacks is the rug pull, where project operators abruptly withdraw liquidity after artificially inflating token value. Existing detection methods primarily rely on reactive on-chain signals and often suffer from temporal data leakage, limiting their real-world reliability. This paper proposes a leakage-aware framework for early rug-pull detection that integrates on-chain behavioral metrics with temporally aligned Open Source Intelligence (OSINT) signals. We construct a hand-labeled dataset of 1,000 token projects, spanning DeFi and non-DeFi settings, with all features extracted strictly prior to any liquidity withdrawal to preserve causal validity. The dataset combines structural on-chain indicators with external attention signals derived from social media activity and search trends. Within this framework, TabPFN is employed as a core modeling component for learning from multimodal tabular data under strict temporal constraints. Experimental results show that the proposed framework achieves strong discriminative performance and improved probability calibration compared to classical baselines, while maintaining low false-negative rates. By framing rug-pull detection as a causal, multimodal forecasting problem, this work emphasizes the necessity of leakage-resilient evaluation and calibrated risk estimation for deployment in blockchain security systems.