The speedy development of generative technologies has resulted in an explosion of deepfake content, with significant risks to digital veracity and public trust. This study offers an effective framework in detecting deepfake images using classical convolutional neural networks (CNNs), particle swarm optimization (PSO), and transfer learning with top-of-the-line pre-trained models. Although initial results with custom CNN and PSO-based optimization showed limited gains, promising performance gains were realized using transfer learning. MobileNetV2 was among the models tested and proved to provide higher accuracy and generalization, making it suitable for low-power and real-time detection of deepfakes. The combination of data augmentation and fine-tuning improved model resilience, solidifying the impact of our method in resource-limited scenarios. The study offers an applied and scalable solution to counteract the influence of synthetic media with effective and precise image-level detection. Evaluation using accuracy, precision, recall, F1-score, and AUC revealed that MobileNetV2 with PSO-based optimization consistently outperformed other deep models. It achieved 77.32% accuracy and 83.75% AUC on the DFRI benchmark, reflecting strong and balanced performance. In contrast, ResNet50, despite PSO tuning, reached only 60.53% precision and 42.59% recall, resulting in a significantly lower F1-score and AUC, indicating poor generalization. EfficientNetB0 demonstrated a high recall of 98.76% but suffered from a low precision of 49.81%, leading to numerous false positives. These comparative outcomes highlight the robustness, stability, and real-world applicability of the MobileNetV2 architecture in deepfake detection, especially in environments with limited computational resources.