AI Provenance Tracker : A Hybrid Deep Learning and Heuristic Approach for Detecting AI-Generated Images

Abstract

Therapidadvancementof generativeartificialintelligencehas significantlyincreasedtheabilityto create highly realistic synthetic images,raisingseriousconcerns regarding digital authenticity, misinformation, and cyber fraud. With the emergence of advanced models such as Generative Adversarial Networks (GANs) and diffusion-based techniques, distinguishing betweenrealandAI- generated images has become increasingly challenging for both humansandtraditionalverification systems.To address this issue,this paper presents an AI Provenance Tracker, an end-to-end hybrid system designed to accurately classify images as real or AI- generated by combining deep learningwithheuristicanalysis. At the core of the proposed system isaResNet50-basedConvolutional Neural Network (CNN) utilizing transferlearning,whichiscapableof extracting complex hierarchical featuressuchasedges,textures,and structuralinconsistenciesfrominput images. The model is trained on a diverse dataset of approximately 40,000 images, consisting of both realandAI-generatedsamples,with appropriate preprocessing techniquesincludingresizing, normalization, and data augmentation.Thenetworkis optimizedusingtheAdamoptimizer and CrossEntropyLoss function, achieving an overall classification accuracy of approximately 90%. In addition to the deep learning model,aheuristicanalysismoduleis incorporated to enhance detection robustness by examining non- learned features such as noise distribution,compressionartifacts, and metadata irregularities, which are often indicative of synthetic content.Theoutputsfromboththe CNN model and the heuristic module are combined using an ensemble strategy based on weightedaveraging,improvingthe system’s reliability and reducing false predictions. Thesystemisimplementedusinga scalable architecture, where the backend is developed using FastAPI for efficient request handling and model inference, while the frontend is built using Next.js and Tailwind CSS to provide an interactive and user-friendlyinterface.Thecomplete workflow enables real-time image analysis, making the system suitable for practical deployment in digital content verification scenarios. Experimental results demonstrate thattheproposedhybridapproach outperforms standalone deep learning models in terms of accuracy, generalization, and robustness. The AI Provenance Tracker offers a practical and scalablesolutiontothegrowing challengeofdetectingAI-generated images, contributing towards enhancing trust and authenticity in digital media ecosystems.

Introduction

The rapid growth of artificial intelligence, especially generative models like GANs and diffusion models, has made it possible to create highly realistic synthetic images. While these technologies provide benefits in creative fields, they also create challenges such as misinformation, digital fraud, and cybercrime. Since AI-generated images are becoming harder to distinguish from real images, reliable authenticity verification has become essential.

Traditional image forensic methods focused on visible errors like lighting issues, unnatural textures, and facial inconsistencies, but modern AI models have reduced these flaws. Deep learning, particularly Convolutional Neural Networks (CNNs), has become a powerful approach because it can identify complex visual patterns and subtle differences between real and generated images. However, existing deep learning models often struggle with unseen data and limited generalization.

To overcome these limitations, the proposed AI Provenance Tracker introduces a hybrid detection system that combines a ResNet50-based CNN model with heuristic analysis. The CNN extracts advanced visual features, while heuristic methods examine additional indicators such as noise patterns, compression artifacts, and metadata inconsistencies. A weighted ensemble approach combines both outputs to improve detection accuracy and reliability.

The system is designed for real-time use with a scalable architecture using FastAPI for backend processing and Next.js for frontend interaction, allowing users to upload images and receive quick authenticity results.

Literature review findings show that deepfake detection has evolved from simple artifact detection to advanced deep learning approaches. Earlier methods used CNN-LSTM models, facial feature analysis, and physical cues like blinking or head movements. Recent research focuses on transfer learning, ensemble models, attention mechanisms, and advanced CNN architectures such as XceptionNet and EfficientNet to improve robustness against new AI-generated content.

Key innovations of the AI Provenance Tracker include:

• Hybrid Detection Model: Combines ResNet50 CNN with heuristic analysis instead of relying only on deep learning.
• Large and Diverse Dataset Training: Uses around 40,000 images from multiple datasets to improve generalization.
• Heuristic Feature Analysis: Detects hidden signs such as noise patterns, compression issues, and metadata errors.
• Ensemble Decision System: Combines multiple detection results for more accurate predictions.
• Real-Time Web-Based Detection: Provides instant image analysis through a user-friendly web platform.
• Scalable Architecture: Designed for applications like social media verification and digital forensics.

Conclusion

This paper presented the AI ProvenanceTracker,ahybridsystem designed to detect whether an image is real or AI-generated. By combining a ResNet50-based deep learning model with heuristic analysis, the proposed approach effectively captures both learned visual patterns and rule-based inconsistencies such as noise and compression artifacts. The system achieved an accuracy of approximately 90%, demonstrating improved performance compared to standalone methods. The use of an ensemble mechanism further enhanced reliability and reduced false predictions. In addition, the scalable architecture built using FastAPI and Next.js enables real- time detection, making the system suitable for practical applications. Overall, the proposed solution provides a robust, efficient, and deployable approach for AI- generated image detection. It contributes toward improving trust, authenticity, and verification in digitalmedia,addressingthegrowing challenges posed by advanced generative technologies.

Copyright

Copyright © 2026 Jyoti Thakur, Suyash Soni, Kritika Markam, Suresh Hemla, Prof. Vishnu Prasad Verma. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.