Advanced Artificial Intelligence-Driven Deepfake Detection: Explainable AI Frameworks for Enhanced Image Authenticity Verification in Digital Media

Abstract

In the digital era, images have become vital evidence in journalism, legal contexts, and social media, where trust in visual content is directly linked to public perception and decision-making. However, the rapid evolution of deepfake technologies, particularly through Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and diffusion models, has introduced unprecedented risks by enabling hyper-realistic image forgeries. These synthetic manipulations threaten information integrity, spread misinformation, and undermine public trust. Addressing this growing concern, the present research investigates the role of Artificial Intelligence (AI) and Explainable AI (XAI) in developing robust and interpretable frameworks for deepfake image detection.

The study utilises a dataset of 1000 samples (real and manipulated images) to design, implement, and evaluate hybrid CNN-attention-based models integrated with explainability techniques such as Grad-CAM, LIME, and SHAP. Preprocessing steps, including normalisation, augmentation, resizing, and noise reduction, were applied to enhance dataset quality and mitigate class imbalance. Feature extraction methods (facial landmarks, texture patterns, and frequency analysis) provided diverse representation for training. Baseline classifiers such as SVM, Random Forest, and simple CNNs were compared against advanced transfer learning models (EfficientNet, ResNet, VGG) to evaluate performance improvements.

The proposed hybrid architecture achieved superior performance across multiple metrics, including accuracy (94.7%), precision (93.8%), recall (95.2%), F1-score (94.5%), and AUC (0.967). Error analysis revealed critical patterns in false positives and false negatives, whilst explainability frameworks successfully highlighted manipulated image regions, increasing user trust and interpretability. Robustness evaluations demonstrated resilience against compression, noise, and adversarial attacks, with cross-dataset testing confirming generalisation capabilities.

Findings indicate that integrating explainability with robust deep learning not only enhances detection accuracy but also ensures transparency for academic, forensic, and societal applications. This research contributes by presenting a scalable, explainable deepfake detection framework that balances accuracy, interpretability, and robustness, emphasising practical applications in fact-checking platforms, forensic analysis, online content moderation, and cybersecurity.