Epileptic seizure detection holds significant importance in medical diagnosis, significantly reducing the workload of physicians. However, the analysis of electroencephalogram (EEG) signals faces challenges of high dimensionality and complexity, making it extremely difficult to extract discriminative features and map them to an appropriate discriminative space for accurate classification. To address this issue, this paper presents a novel seizure detection approach using Deep Adversarial Metric Learning (DAML), this method incorporates a structured three-phase optimization process designed to progressively enhance the performance of the feature extraction component. In the first stage, triplet loss pre-trains feature extraction, refining embedding space and enhancing feature discriminability. In the second stage, adversarial metric learning is employed to generate challenging negative samples, refining inter-class margins and improving system resilience. Finally, the generator and metric learning modules are removed, retaining only the optimized feature embeddings, and a K-Nearest Neighbors (KNN) classifier is introduced for classification. Experimental results demonstrate that the proposed method achieves 96.68% accuracy, 92.26% sensitivity, and 98.95% specificity on the CHB-MIT scalp EEG dataset, which includes data from 24 patients. The findings confirm that our method enhances seizure detection accuracy, suggesting its strong potential for clinical use.