Abstract Objectives: This study aims to develop an automated and scalable deep learning framework for detecting vulnerabilities in Ethereum smart contracts by solving the limitations of traditional manual auditing and static analysis methods which mainly suffer from high false positives and limited pattern recognition capability. Method: A novel bytecode-to-image transformation pipeline has been designed to convert Ethereum Virtual Machine (EVM) bytecode into standardized 128×128 grayscale images. Multiple deep learning architectures which were trained and evaluated on a dataset of over 12,000 annotated smart contracts across eight vulnerability classes. SMOTE-based class balancing and multi-class classification techniques which were applied to improve model robustness. Findings: Baseline CNN and MobileNetV3-Large has achieved accuracies of 0.75 and 0.39 respectively while EfficientNet-B7 improved performance to 0.77 accuracy. The proposed EfficientNet-B7 + CBAM model has achieved the highest performance with 0.80–0.81 accuracy, 0.80 precision, 0.80 recall and a macro F1-score of 0.79 by showing the performance of attention-guided feature refinement. Novelty: The study introduces a language-independent visual representation of smart contract bytecode combined with attention-enhanced deep learning by giving a scalable and quite a good alternative to traditional vulnerability detection approaches in blockchain security. Keywords: Vulnerability Detection, Smart Contracts, Deep Learning, EfficientNet-B7, Blockchain Security
Mahale et al. (Fri,) studied this question.