Facial forgery technology presents significant challenges to the integrity and security of digital content. This study explores the enhancement of deepfake detection models through adversarial training, emphasizing the resilience of these models against maliciously engineered inputs. We assessed the performance of both baseline and adversarially trained convolutional neural network (CNN) and VGG16 models, using clean and adversarial test datasets. The results indicate that while baseline models experience notable performance declines under adversarial attacks, adversarially trained models exhibit considerable improvements in resilience, achieving a 97% accuracy on adversarial samples without significantly sacrificing accuracy on clean data. Beyond static evaluations, we developed a real-time deepfake detection system that demonstrates practical applications, including security in live environments. This system illustrates the efficacy of adversarially trained models, showcasing robust detection capabilities in real-world scenarios. Our findings highlight the effectiveness of adversarial training in enhancing the robustness of deepfake detection systems against manipulation while preserving operational efficiency. This research offers valuable insights into creating secure and reliable machine learning models and underscores the critical role of adversarial robustness in the advancing field of digital security. Future research will focus on refining these techniques and investigating their applications across various contexts and modalities.