The ease of access to content on social media, along with modern tools and low-cost computing infrastructure, has led to the creation of deepfakes that disseminate deception and forgeries. This quick advancement may bring panic and instability because anyone can simply create deception using these technologies. In the modern era of social media, it’s vital to have a reliable system for distinguishing between genuine and fraudulent content. This study offers a novel lightweight end-to-end deep neural network for categorizing deep fake images using spatial and self-attention methods. The attention mechanism learns contextual information from convolutional neural network features and emphasizes discriminative aspects to grasp the relationship between authentic and deepfake images. The proposed model performed well on the OpenForensics dataset, with a 91% accuracy. Furthermore, to prove the generalisability of the proposed network, a test set from the DeepFake Detection dataset is utilized to examine the proposed approach trained on the OpenForensics dataset. The model’s efficiency (0.34M parameters) and cross-dataset accuracy (95% on DFDC) make it viable option for real-world deployment, such as social media moderation or forensic analysis. The Github link of the source code is: https://github.com/KarnatiMOHAN/DeepFake-Model .