Combating Market Manipulation: A WassersteinGAN GP Augmented, Focal Loss TabNet for Cryptocurrency Pump and Dump Detection

Pump-and-dump (P&D) manipulation is one of the most persistent forms of cyber-enabled fraud in cryptocurrency markets, and detecting it automatically is made difficult by the extreme scarcity of confirmed P&D events relative to normal trading activity. This study develops an augmentation strategy that integrates a fully specified Wasserstein GAN with gradient penalty, and an alpha-balanced focal-loss TabNet classifier for transaction-level P&D detection. A WGAN-GP generator is trained independently within each cross-validation fold on the training partition only, contributing 30,000 synthetic minority rows per fold, combined with a bounded thirty-fold replication of genuine minority transactions to reach a documented augmented imbalance ratio of approximately 8.9 majority rows to one minority row. On the primary 25-second resolution, the full proposed model attained a precision of 100.0%, recall of 85.7%, and F1-score of 92.3% on a stratified held-out test split, and a cross-validated mean F1-score of 0.875 (precision 0.878, recall 0.874) under 5-fold stratified cross-validation with fold-wise generator retraining. A five-configuration ablation isolates the marginal contribution of WGAN-GP synthesis (ΔF1 +0.019 over a class-weighted baseline), hard-negative mining (+0.006), and focal loss (+0.038), demonstrating that the generative component contributes meaningfully but that the largest single improvement is attributable to the loss function rather than to the generator. Applying the 25-second-tuned hyperparameter configuration without re-tuning to the 15 and 5 second aggregations yields cross-validated F1-scores of 0.840 and 0.766 respectively, quantifying for the first time on this corpus the cost of resolution under-tuning in a GAN-augmented pipeline.