Privacy-Preserving Blockchain Transactions Synthesis with Diffusion Generative Models

Blockchain transactions are increasingly vulnerable to sophisticated attacks, necessitating detection mechanisms that safeguard user privacy. Traditional methods, such as CTGAN and PATE-GAN, often sacrifice privacy to enhance detection accuracy, or compromise detection capabilities to maintain privacy, struggling to effectively balance both. This paper introduces BTGM (Blockchain Transactions Generative Models), an advanced approach that ingeniously combines differential privacy with a two-stage diffusion process, and incorporates Row Modeling based on VGM (Variable Gaussian Mixture) encoding. This hybrid model not only ensures the robust anonymization of transaction data through differential privacy but also effectively learns the distribution of data features via its innovative use of Row Modeling and diffusion processes. Validated on the Elliptic++ dataset, BTGM achieves 0.96 accuracy and 0.86 PRAUC, outperforming SOTA models like CTGAN and STaSy. It effectively thwarts membership inference attacks, maintaining a strong balance between privacy protection and attack detection.