With the accelerated pace of artificial intelligence, deepfake technology has advanced and threatens the authenticity and security of digital media. Deepfakes use Artificial Intelligence methods to edit videos such that it is difficult to distinguish between authentic and manipulated content. This paper delves into detecting deepfakes through two novel deep learning-based methods: Vision Transformers (ViTs) and MesoNet. Unlike standard CNNs (convolutional neural networks), ViTs (Vision Transformers) utilize self-attention mechanisms to understand the long-range dependencies among video frames, demonstrating remarkable effectiveness in identifying subtle discrepancies commonly found in deepfake videos. In contrast, MesoNet is tailored to detect facial manipulations, targeting forgery methods like Face2Face and Deepfake. The method is strictly tested on benchmark datasets and proves highly accurate and resilient. The ViT model reached a maximum accuracy of 91. 78% in FaceForensics++ and 92. 81% in Celeb-DF (V2), while MesoNet reached 98. 16% accuracy in Celeb-DF (V2), highlighting their strength in recognizing manipulated videos. The results highlight the strengths of both models, with ViTs being superior at modeling global dependencies and MesoNet having excellent facial forgery detection. This study provides an important step toward creating secure deepfake detection methods to address increasing alarms on misinformation, identity deception, and digital dishonesty. The combination of ViTs and MesoNet is a promising approach to making deepfake detection systems more reliable and secure.