AI-Based Deepfake Image Detection: Robust and Explainable Approaches for Ensuring Digital Image Integrity

Abstract

In an increasingly digital society, the authenticity of images plays a decisive role in shaping public trust across journalism, law enforcement, forensic analysis, and social media communication. The emergence of deep fakes, enabled by advanced generative models such as Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and diffusion-based frameworks, has created a powerful threat to digital image integrity. These manipulations not only enable the spread of misinformation but also compromise legal evidence and erode confidence in digital ecosystems. Detecting such sophisticated forgeries requires robust artificial intelligence (AI) methods that are not only accurate but also explainable and transparent.

This research presents a comprehensive study on robust and explainable deep fake image detection, using a curated dataset of 1000 images (real and manipulated). A multi-phase research design was employed, beginning with dataset preparation, preprocessing, and feature extraction, followed by the development of advanced AI architectures. The proposed hybrid CNN-attention model integrates transfer learning (EfficientNet, ResNet, VGG) with an explainability framework powered by Grad-CAM, LIME, and SHAP to provide interpretable outputs for end-users. The dataset underwent rigorous preprocessing (normalization, augmentation, resizing, denoising), ensuring balanced representation and reduced bias.

Baseline machine learning models (SVM, Random Forest, simple CNNs) were benchmarked against the proposed architectures to establish performance improvements. The evaluation metrics included accuracy, precision, recall, F1-score, AUC, and newly introduced explainability scores to measure interpretability. The results indicate that the hybrid CNN-attention model consistently outperformed baselines, achieving 94.7% accuracy and delivering visual explanations that highlighted manipulated image regions, thereby enhancing user trust. Security and robustness evaluations demonstrated the resilience of the proposed model under conditions of compression artifacts, noise perturbations, and adversarial attacks.