Blockchain fraud detection confronts escalating challenges in Ethereum ecosystems where evolving fraud patterns result in billions of dollars in annual losses, yet existing ap proaches fail to adapt without catastrophic forgetting while pro viding no theoretical guarantees for financial system deployment. This paper presents AHGT-DFD, a framework integrating four ML techniques automated pattern discovery through Dirichlet process priors, hierarchical variational learning, heterogeneous graph transformers, and continual learning mechanisms into a theoretically grounded system. We provide rigorous theoret ical foundations including PAC-learning bounds guaranteeing generalization performance within 6.21% of empirical error with 95% confidence, polynomial-time convergence analysis, and certified robustness against adversarial perturbations. Compre hensive evaluation on real-world Ethereum datasets demonstrates 4.62 percentage point F1-score improvement for Ponzi detection (95.58% vs. 92.15% best baseline) and 4.94 percentage point improvement for phishing detection (97.41% vs. 94.47% best baseline). The framework maintains 94.2% performance reten tion on evolving patterns compared to 78.6% for conventional approaches, while supporting real-time processing with sub-15ms inference latency suitable for production blockchain security systems. This integration with theoretical guarantees provides a foundational advance for dependable fraud detection in adver sarial financial environments.

Non-fungible token (NFT) markets present a dual analytical challenge: integrating heterogeneous data modalities (high-dimensional visual features and discrete transaction sequences) while preserving privacy for sensitive wallet addresses and trading strategies. Current approaches analyze visual attributes or transaction patterns in isolation, missing critical value drivers from cross-modal interactions. Meanwhile, existing multimodal techniques lack formal privacy guarantees, exposing participants to inference attacks. This article introduces PrivaMod, a privacy-preserving Bayesian framework that addresses these limitations through uncertainty-aware multimodal fusion. Our approach implements precision-weighted Bayesian fusion that dynamically adjusts modality contributions based on quantified uncertainty levels, while integrating Rényi Differential Privacy throughout the pipeline via calibrated noise injection and adaptive gradient clipping. Evaluated on 167,492 CryptoPunk transactions, PrivaMod achieves a market efficiency score of 0.874 and R 2 of 0.912, outperforming existing methods by 13.4% through superior cross-modal integration while maintaining strong privacy guarantees ( \(\varepsilon\) = 0.08, \(\delta\) = 1e-5) with membership inference attack success rates near random guessing (53.4%). The system demonstrates that privacy-preserving techniques can enhance rather than compromise analytical performance, establishing a foundation for responsible market analysis. To ensure reproducibility, we release our code, preprocessed datasets, and model checkpoints with detailed documentation and scripts to replicate all experiments. PrivaMod is available at https://github.com/kvjunior/PrivaMod/blob/main/README.md .