Eliminating Sandwich Attacks with the Help of Game Theory

TL;DR

This paper introduces a game-theoretic model to analyze sandwich attacks in Ethereum's AMMs and proposes an adaptive algorithm for traders to set slippage tolerance, significantly reducing attack costs.

Contribution

It develops the sandwich game for analytical insights and presents an adaptive slippage algorithm that outperforms existing fixed auto-slippage methods.

Findings

Most transactions can avoid sandwich attacks with low risk of failure.

The proposed algorithm reduces attack costs by over 100 times compared to fixed auto-slippage.

Adaptive slippage setting outperforms constant auto-slippage in all tests.

Abstract

Predatory trading bots lurking in Ethereum's mempool present invisible taxation of traders on automated market makers (AMMs). AMM traders specify a slippage tolerance to indicate the maximum price movement they are willing to accept. This way, traders avoid automatic transaction failure in case of small price movements before their trade request executes. However, while a too-small slippage tolerance may lead to trade failures, a too-large slippage tolerance allows predatory trading bots to profit from sandwich attacks. These bots can extract the difference between the slippage tolerance and the actual price movement as profit. In this work, we introduce the sandwich game to analyze sandwich attacks analytically from both the attacker and victim perspectives. Moreover, we provide a simple and highly effective algorithm that traders can use to set the slippage tolerance. We unveil that most broadcasted transactions can avoid sandwich attacks while simultaneously only experiencing a low risk of transaction failure. Thereby, we demonstrate that a constant auto-slippage cannot adjust to varying trade sizes and pool characteristics. Our algorithm outperforms the constant auto-slippage suggested by the biggest AMM, Uniswap, in all performed tests. Specifically, our algorithm repeatedly demonstrates a cost reduction exceeding a factor of 100.