ScanAT: Identification of Bytecode-Only Smart Contracts With Multiple Attribute Tags

Abstract

Smart contracts on blockchain systems implement business logic and directly handle important assets. Although smart contracts play these critical roles, it is hard for users interacting with the system to understand the real behavior of the deployed bytecodes of smart contracts. The quirks of smart contracts, such as code reuse and limited unique datasets, make it challenging to recognize the functional details of smart contracts. In this paper, we propose a new method for characterizing bytecode-only smart contracts by automatically assigning multiple attribute tags. Using a deep learning approach, our system, the ScanAT, extracts attribute tags from the source code and metadata of known smart contracts and trains their bytecode with the attribute tags. The ScanAT then infers attribute tags from the bytecode of smart contracts alone. Our experiments show that ScanAT can achieve 81% accuracy in predicting attribute tags, using convolutional neural networks and a customized autoencoder.