Extending the generative data augmentation method proposed in Chap. 3, this chapter addresses the challenge posed by the substantial overlap between fraudulent and legitimate samples. This chapter proposes a GAN-based Hybrid Sampling (GANHS) framework [33], which integrates a Behavior-Boundary Aware Generative Adversarial Network (BBAGAN) for data augmentation with an Adaptive Neighborhood Cleaning Strategy (ANCS). This approach simultaneously generates high-quality fraudulent samples while eliminating ambiguous legitimate samples. It is particularly effective in detecting fraudulent activities characterized by weak discriminatory signals and strong concealment, such as micro-payment abuse and behavioral-camouflage transactions [14].