Abstract—The rapid advancement of deepfake technology has raised significant concerns regarding the authenticity of digital facial images, posing threats to privacy, security, and trust in media. This study presents a robust and explainable framework for deepfake image detection by leveraging Convolutional Neural Networks (CNNs) enhanced with Convolutional Block Attention Modules (CBAM), which help the model focus on key tampered regions. The detection is formulated as a binary classification task between real and fake facial images. A two-stage preprocessing pipeline is proposed, combining Error Level Analysis (ELA) and Structured Forest Edge Detection (SFED) to amplify forgery traces. Five different CNN architectures are employed in parallel; each integrated with CBAM to enhance discriminative feature learning. To improve interpretability, Gradient-Weighted Class Activation Mapping (Grad-CAM) highlights the most influential image regions that contribute to the model’s decisions. A stacking ensemble model using extreme gradient boosting (XGBoost) aggregates the individual predictions for improved generalization. The proposed system is evaluated on a large benchmark dataset of 140K FAKE AND REAL FACES, achieving 97.25% accuracy, 96.88% F1-score, and 99.68% Area Under Curve (AUC), demonstrating both high performance and interpretability.