Optimistic rollup has emerged as a promising Layer 2 (L2) scaling solution for blockchain; however, its existing protocols are vulnerable to front/back-running activities, where an opportunistic rollup operator can strategically alter the transactions' order to create an arbitrage opportunity. Specifically, in the limited edition ERC-721 standardized non-fungible tokens (NFTs), the re-ordering of transactions introduces a lucrative threat landscape due to its scarcity-driven pricing and market volatility. In this work, we introduce PAROLE, a novel attack technique on optimistic rollup systems, where an adversarial aggregator re-orders the NFT transactions in an optimal way, leveraging model-free deep reinforcement learning (DRL) to maximize the balance of a target account. We create our own NFT called the 'PAROLE Token' (PT) and deploy it in the OpenSea marketplace via Optimism Goerli to validate the attack impact. Furthermore, we collect NFT snapshots from rollup mainchains to analyze the impact in real-world NFT marketplaces.
