Do Classifiers and Compression Agree? A Leakage-Aware Comparison of Sybil Detection on Ethereum

Authors

Abstract

Ethereum projects attract new users with airdrops–public giveaways of crypto tokens. This is an opportunity for Sybil bots–wallets operated by the same entity to abuse the rewards system. There are established Sybil detection methods, but every detector method family has its price: classifiers have high labeling and retraining costs; graph-based methods can be evaded by rerouting transactions through bridges or centralized exchanges; and compression-based techniques are training-free but costly to scale. Picking the correct method may be difficult because these methods have been studied in isolation. In this paper, we ask the following questions: which family of techniques is best for leakage-aware classification; which family is best for training-free Sybil candidate discovery; and do the two types of audit regimes–closed-set classification and open-world discovery–detect the same Sybils? Our approach is to combine tabular, sequential, and compression-based techniques in one evaluation scheme: our Transaction Grammar decodes the rhythm, execution structure, and intent of 14161 Ethereum wallets participating in the Hop airdrop, with all class-exclusive counterparties filtered out via the Blind-Spot protocol. We contribute a comparison of tabular (XGBoost, SVMs), sequential (Transformers, BiLSTMs), and compression-based (NCD) Sybil detectors in one leakage-controlled setting, and a two-stage pipeline that discovers more of a Sybil campaign than either technique used separately.

Additional Files

Published

2026-10-26

Issue

Section

Applied Informatics