GenPix: A Diverse Dataset for Fake Image Detection

Deep Learning-Based Fake Image Detection Using Transfer Learning: A Systematic Review

The rapid advancement of artificial intelligence (AI) has significantly changed the way digital visual content is created, enabling the generation of highly realistic synthetic images and videos. While these technologies support many beneficial applications, they have also facilitated the creation of manipulated visual content, commonly known as deepfakes, which pose serious challenges to information authenticity, public trust, cybersecurity, and digital forensic investigations. As image manipulation techniques continue to evolve through advanced models such as Generative Adversarial Networks (GANs) and diffusion-based frameworks, conventional detection methods relying on handcrafted features have become increasingly inadequate. In response, deep learning approaches integrated with transfer learning have emerged as effective solutions due to their ability to leverage pre-trained models for extracting robust and discriminative features, even when limited training data are available. This review presents a comprehensive analysis of recent deep learning and transfer learning techniques for fake image detection. It examines widely adopted convolutional neural network (CNN) architectures, benchmark datasets, evaluation metrics, and current research developments. Furthermore, the paper provides a comparative assessment of existing methods by highlighting their strengths, limitations, and performance characteristics. Finally, it identifies major research challenges and outlines future directions for developing robust, scalable, and generalizable fake image detection systems capable of addressing the growing threats posed by AI-generated visual content in cyberspace

Training-Free Reconstruction-Based AI-Generated Image Detectors Are Inherently Vulnerable to Adversarial Examples

The impressive visual quality and ubiquity of AI-generated images call for reliable and robust detection methods. Reconstruction-based detectors have emerged as a promising direction for transparent and training-free identification of synthetic images. However, due to their fundamentally different mode of operation (compared to standard, classifier-based methods), little is known about their adversarial robustness. In this work, we propose two novel attack methods targeted at detectors that leverage autoencoder reconstruction error. We find that by constructing imperceptible adversarial examples, the distance between original and reconstruction can be artificially increased, causing fake images to be wrongly classified as real. Our evaluation including images from three state-of-the-art generators and three detectors demonstrates that detection performance is significantly decreased, even if attacked images additionally undergo real-world degradations. Critically, our adversarial examples naturally transfer across detectors, as they all share the same principle, pointing towards an inherent vulnerability of reconstruction-based detectors.

Generalized Synthetic Image Detection with Enhanced RGB-Noise Representation Learning

The rapid advancement of large-scale generative models has accelerated the spread of highly deceptive AI-generated images, making generalized synthetic image detection a critical imperative. Existing forensic networks often struggle with cross-model generalization and realworld degradations due to their reliance on single-domain representations and conventional binary classification optimization. To overcome these limitations, we propose RNSIDNet, a novel forensic framework that achieves robust detection through enhanced RGB-Noise representation learning. Specifically, our method employs a dual-branch architecture where global RGB semantics, extracted by an attention-refined CLIP backbone, dynamically modulate highfrequency noise artifacts captured by Bayar convolutions via a Feature-wise Linear Modulation (FiLM) module. To further enhance the learned representations, we design a Hard Sample-aware Contrastive Learning (HSCL) strategy. By explicitly penalizing challenging training samples, HSCL reshapes the latent feature space to maximize the discriminative margin between pristine and synthetic domains. Extensive experiments across eight public benchmark datasets verify that our model achieves state-of-the-art performance, delivering superior generalization ability, robustness, and computational efficiency. Code and dataset will be publicly available on https://github.com/multimediaFor/RNSIDNet.

Enhancing Cross-Dataset Generalization in Image Forgery Detection via Mixed-Domain Training and Adaptive Thresholding

Image forgery detection is an important part of digital forensics especially as more and more sophisticated image manipulating tools become more and more accessible. Despite the fact that deep learning models have shown high accuracy on benchmark datasets, their performance on heterogeneous datasets is a major challenge because of the domain shift and changes in manipulation methods. The paper suggests a hybrid training system and adaptive thresholding to improve the generalization of cross-datasets in image forgery detection. A model trained on CASIA dataset performed well on in-domain data but has a significant decrease when tested on unseen data with only a 48.5% accuracy on the COVERAGE dataset. To overcome this weakness, a mixed-domain learning approach that used samples across various datasets such as CASIA and COVERAGE was proposed in order to enhance feature robustness. The suggested solution was tested using three benchmark data sets, including CASIA, Columbia and COVERAGE. The experimental results show high accuracy of CASIA (approximately 95%), Columbia (approximately 95%), and considerably better accuracy on COVERAGE at 91.4%. The model also obtained high AUC-ROC value of 0.994 which implies that there was great class separability. Further analysis based on the precision, recall and F1-score validates balanced and consistent performance. Moreover, qualitative analysis in terms of score distribution and prediction visualization depicts obvious discrimination between genuine and manipulated photos. The results indicate that mixed-domain training is a practical approach that can reduce dataset bias and increase generalization, which is why the proposed framework can be adopted in the real-world forensic use. The work offers a complete solution to domain adaptation issue in the image forgery detection.

Dual-stream framework for real/fake classification and source attribution of AI-generated images using spatial and frequency features

A Deep Generative Framework for Realistic Colorization of Grayscale Astronomical Images

Astronomical images captured using Space based or Ground based telescopes are majorly in grayscale, hiding the minute details of the intricate structures. Our paper presents a novel idea of the implementation of a Generative Adversarial Network (GAN) - based framework for astronomical image colorization. Our proposed model utilizes a U-Net-based generator and a PatchGAN discriminator to learn realistic color schemes from the grayscale input. The system was trained on the Hubble Space Telescope data set collected from NASA and ESA archives, comprising 673 high-resolution images, with an additional 99 images reserved for testing. The images are restricted to a resolution of 512 x 512 pixels due to computational limitations. Quantitative evaluation demonstrates that our proposed framework achieves an average Peak Signal-to-Noise Ratio (PSNR) of 31.46 dB, Structural Similarity Index Measure (SSIM) of 0.92, and Signal-to-Noise Ratio (SNR) of 26.85 dB which outperforms traditional approaches. The experimental results confirm that the model produces perceptually accurate, visually coherent, and scientifically meaningful and colorized outputs which contribute to enhanced visualization and analysis of astronomical data.