GAN-TL domain-adaptive augmentation for enhanced transfer learning

Transfer learning offers a practical solution for label-scarce target domains, but its effectiveness diminishes when source and target data follow different distributions. To address this challenge, we present GAN-TL, a GAN-augmented transfer learning framework that generates target-style samples to expand target-relevant variability and couples this synthesis with adversarial feature alignment. The task model is trained on mixed real and synthetic batches, enabling improved cross-domain generalization under limited target supervision. Experiments on three benchmarks (Datasets A–C) confirm consistent improvements over standard transfer learning. On Dataset A, GAN-TL achieves 87.4% accuracy compared with 83.5% for conventional fine-tuning, and increases macro-F1 from 83.4% to 87.4%. For the regression benchmark Dataset B, mean squared error decreases from 9.72 to 7.89 while $$R^2$$ improves from 0.83 to 0.88. On the unseen-domain Dataset C, GAN-TL raises accuracy to 82.3%, reduces domain adaptation loss from 0.25 to 0.19, and lowers generalization error from 0.22 to 0.18. Robustness evaluations further show improved resistance to adversarial perturbations, with FGSM attack success reduced to 39.2% (vs. 48.3%) and average DeepFool perturbation lowered to 0.009 (vs. 0.012). Despite the added GAN branch, training converges faster in epochs, reaching stability at 26 epochs with a final loss of 0.53 (vs. 29 epochs and 0.62 for standard transfer learning). Overall, the results demonstrate that target-style synthesis combined with principled alignment yields a stable and robust transfer pipeline under pronounced domain shift.