AI-Driven Digital Twin Framework for Predictive Risk Assessment of Upgradeable Smart Contracts in Blockchain-Based Metaverse Applications

The increasing adoption of blockchain technology in metaverse applications has accelerated the use of upgradeable smart contracts to support maintainability and functional evolution without redeployment. However, the proxy mechanisms enabling upgradeability also introduce additional security risks related to delegation, administrative control, and storage management. Existing studies primarily focus on proxy smart contract detection and provide limited support for predictive security assessment under different operational conditions. This study proposes an AI-Driven Digital Twin framework for predictive smart contract security analysis that integrates leakage-safe machine learning with scenario-based Digital Twin simulation. A dataset consisting of 377 Ethereum smart contracts, including 17 proxy and 360 non-proxy contracts, was analyzed using static structural features extracted from Solidity source code. Logistic Regression and Random Forest classifiers were evaluated using leakage-safe grouped cross-validation to prevent structurally similar contracts from appearing simultaneously in the training and testing sets. The Digital Twin representation was subsequently used to estimate heuristic security risks under four operational scenarios, namely Baseline, Upgrade, Adversarial Caller, and High Load. The experimental results show that Logistic Regression achieved the best overall performance with 97.61% accuracy, 98.75% balanced accuracy, perfect recall, an F1-score of 0.7907, a Matthews Correlation Coefficient of 0.7984, a ROC-AUC of 0.9980, and a PR-AUC of 0.9647. Feature importance analysis identified EIP-1967-related indicators as the strongest predictors of proxy contracts, while the Digital Twin simulation consistently assigned higher risk scores to proxy contracts across all evaluated scenarios, with the Upgrade scenario producing the highest estimated risk. These findings demonstrate that the proposed framework extends conventional proxy detection by providing leakage-safe classification together with scenario-dependent predictive security assessment, supporting pre-deployment decision making for secure blockchain-based metaverse applications.