AI-GENERATED IMAGE DETECTION USING SIXCHANNEL MULTI FORENSIC CONVOLUTION NUERAL NETWORKS (CNN) FUSION

Modern generative artificial intelligence (AI) can generate highly realistic synthetic images, making current detection system that rely on standard RGB input tends to miss the underlying generative artifacts. This research develops a framework based on Convolutional Neural Network (CNN) that aims to distinguish between original digital images and artificial media. This work proposes a multi-modal, six channel forensic fusion architectures that combines standard visual data with advanced digital forensic features to reveal hidden generative artifacts. The proposed method utilizes three specific forensic layers, Error Level Analysis (ELA) to find inconsistencies in image compression, Fast Fourier Transform (FFT) to detect abnormal frequency patterns, and Laplacian High-Pass Filtering (HPF) to extract high-frequency background noise as Photo Response Non-Uniformity (PRNU) alternative. The first layer of three baseline models which are MobileNetV3-small, MobileNetV2, and ResNet50 were modified to directly process the combined six-channel input. The framework was trained, validated and tested based on balanced subset of 16,000 images from the CIFake dataset. The performance of all models was measured using accuracy, precision, False Acceptance Rate (FAR), False Rejection Rate (FRR), F1-score, and overall computational cost. Implementing these forensic features significantly improved the accuracy of lightweight models. As a result, MobileNetV2 accuracy jumped from 77.62% using only standard RGB to 94.94% with the new six channel input, while MobileNetV3-small achieved the highest overall test accuracy of 95.69%. This multi feature approach can be used in real-world applications like digital forensics, automated content moderation, and cybersecurity, where verifying image authenticity is critical. Overall, the findings indicate that combining traditional method with forensic features with CNN-based model can improve detection accuracy, even with lightweight models