As blockchain ecosystems like Base (an Ethereum Layer 2 solution) grow rapidly, token creation has become increasingly common. However, this growth has also led to malicious tokens designed for fraud—stealing funds, running scams, or exploiting smart contract vulnerabilities. These malicious tokens are digital assets created to deceive users, investors, and network participants through hidden harmful code. Early detection of such tokens is vital for user protection and ecosystem integrity, though this proves challenging due to blockchain's pseudonymous nature and attackers' sophisticated obfuscation tactics. Malicious tokens typically show distinct patterns of suspicious behavior, including unusual transaction flows, concentrated ownership, vulnerable contract code, and irregular token economics. A data-driven approach to analyzing these characteristics is essential for effective detection. A machine learning system using on-chain transactions and token metadata offers a solution to this challenge. This approach starts by gathering and processing blockchain data to identify signs of malicious activity, particularly focusing on ownership patterns and transaction behaviors. These indicators feed into a predictive model—whether Random Forest, Gradient Boosting, or Graph Neural Network (GNN)—that learns to distinguish between suspicious and legitimate tokens. Adding anomaly detection capabilities helps catch new, previously unseen attack patterns, creating a comprehensive defense for the network and its users. If you are unfamiliar with the basic concepts in Crypto such as tokens and wallets, please start with our blog post "[Blockchain 101](https://doc.cryptopond.xyz/docs/blockchain-101-for-data-scientists)". Otherwise, let's dive in! Note: To qualify for competition points, please upload **both your model predictions and your code**. [Competition Webpage](https://cryptopond.xyz/modelfactory/detail/223794)