The ISCTürkiye 2025 paper presents an Advanced Variational Autoencoder framework for classifying Ethereum wallets associated with Lazarus-linked activity. The model uses 116 behavioral indicators covering graph topology, temporal dynamics, transaction flows, service-specific behavior, and cross-chain bridge interactions. It was trained on 15,260 Ethereum wallets, split between 7,603 Lazarus-labeled wallets and 7,657 non-Lazarus wallets that include licit users, mixers, nested services, bridges, and other illicit non-Lazarus activity. The reported results are 98.998% accuracy, 99.128% precision, 98.862% recall, and 99.947% AUC, with self-attention, residual connections, and a combined reconstruction/classification objective. The work matters for blockchain forensics because it aims to distinguish Lazarus-style wallet behavior from other illicit and legitimate cryptocurrency patterns.