<p>Background: Accurate classification of epileptic seizures into focal and non-focal types is essential for effective treatment planning, yet remains a challenging task due to the subjective nature of EEG interpretation and limited availability of labeled data. Traditional approaches based on feature extraction and conventional deep learning models often struggle with generalization when faced with class imbalance and small datasets. New Method: We introduce a novel deep learning framework combining a class-specific Generative Adversarial Network (GAN) and a 1D Convolutional Neural Network (CNN). Each generator in the GAN is specialized to synthesize either focal or non-focal EEG signals, learning the statistical and physiological features unique to its class. Input samples were downscaled by splitting original 10,240-sample EEG vectors into 5,120-sample segments, resulting in a more stable and efficient GAN structure. The generated synthetic signals are then used alongside real samples to train a CNN classifier. Results: The class-specific GAN achieved fidelity scores of 99.68% for focal and 99.43% for non-focal signals. The CNN classifier trained on the augmented dataset achieved 99.56% accuracy, 99.56% sensitivity, 99.56% specificity, 99.57% precision, and 99.56% F1-score, with a Cohen’s Kappa of 99.13% and only 0.87% false positive rate. Comparison with Existing Methods: Our method outperformed prior approaches including-SVM (83%), autoencoder-based models (~ 93%), and GAN variants such as WGAN-GP (91.73%), achieving new state-of-the-art performance on the Bern-Barcelona EEG dataset. Conclusions: The proposed class-specific GAN and CNN framework effectively addresses data scarcity, improves signal realism, and enables near-perfect classification of focal vs. non-focal EEG, advancing AI-driven epilepsy diagnostics.</p>