AI-generated artwork detection using self-distilled transformers with global-local feature learning and Grad-CAM interpretability

Sep 30, 2026

Upload by Jacob

AI-generated artwork detection using self-distilled transformers with global–local feature learning and Grad-CAM interpretability presents a deep learning framework to distinguish human-created art from highly realistic machine-generated images. The approach leverages the DINO v2 self-distilled transformer to extract discriminative global structures alongside fine-grained visual cues, trained and evaluated on a balanced dataset of real and AI art. Results reach 99.01% accuracy, 95.29% precision, 94.58% recall, 94.93% F1-score, and 99% AUC. Interpretability and statistical validation support dependable, transparent predictions.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages13

Reading Time33 min

AccessFree to download

Introduction

Background of AI-generated art and detection needs

Related Work

Limitations of traditional feature extraction and statistical methods

Methodology

Dataset, preprocessing, baselines, and DINO v2 framework

Training, evaluation, and interpretability

Results and Validation

Performance metrics and statistical p-value validation

[www. nature.com/scientificreports](www. nature.com/scientificreports) OPEN AI-generated artwork detection using self-distilled transformers with global–local feature learning and Grad-CAM interpretability Wang Yinghua1􀀍, Li Linyan1, Ma Wenjuan1 & Zhang Yunzhe2 This study presents a strong framework for the detection of artificial intelligence-generated artwork using digital imaging and deep learning–based transformers models, which helps in art community to distinguish the authenticity of human-created art from highly fascinated machine-generated content. Art holds reflection of deep cultural, historical, and social significance, however, due to rapid advancements in artificial intelligence, particularly in generative adversarial networks and diffusion models, have enabled the production of visually creating artworks that blur the boundaries between originality and synthesis. Traditional methods relying on conventional features and statistical analysis are increasingly lower performance against such advance transformation, highlighting the need for more advanced detection mechanisms. To address this challenge, the proposed approach employs Distillation with No Labels (DINO) v2, a self-distilled transformer model that excels in extracting discriminative features by capturing both global structures and fine-grained visual cues. The model was trained and evaluated on a balanced dataset of real and AI-generated art images, with results benchmarked against strong baselines. Experimental findings demonstrate that the proposed framework achieves 99.01% accuracy, 95.29% precision, 94.58% recall, 94.93% F1-score, and an AUC of 99%, outperforming all baselines with superior generalization. Furthermore, interpretability methods along with statistical validation based on log of p-values, confirmed that predictions are both dependable and transparent. Keywords Artificial intelligence, Computer vision, Feature extraction, Art detection, Content authenticity Art has always shown reflection of human culture, creativity, and identity based on ancient cave paintings to modern digital illustrations. It captures not only aesthetic beauty but also increases the values, emotions, and narratives of societies across centuries1. With its ability to exceed geographical and temporal boundaries, art serves as a universal language that communicates human experience in ways words often cannot2. In today’s digital era, however, the landscape of artistic creation has been gradually transformed, targeting the very actual authenticity, originality, and authorship3. The emergence of artificial intelligence (AI) has brought new aspects to the creation of art. More sophisticated models like Generative Adversarial Networks (GANs) and diffusion-based models have now been able to generate artworks that are highly realistic and can have a variety of styles4. These innovations provide new opportunities to be creative, on the other hand, they are very problematic to the art authentication, cultural preservation and intellectual property5. The conventional authentication methods, which were once effective to detect any manual forgery, are no longer effective in detecting patterns of irregularities being built into AI-generated images6. Art trend is becoming increasingly day-by-day because of the introduction of numerous AI-tools to create images. This widening divide reaffirms the urgent necessity of smart, automatic systems of detection that will be able to distinguish between real artworks created by man and the machine-generated fakes7. The previous approaches were based on the intensive use of traditional feature extraction8 and the statistical analysis of image textures, color histograms, and brushstroke patterns9. Although these methods offered useful information, their low flexibility to investigate more complicated generative processes and generalization over a variety of datasets greatly weakened their performance10. As the technologies of AIs-

FAQs

What problem does the study address in AI art detection?

It targets the challenge of distinguishing authentic human-created artworks from visually compelling AI-generated images that blur originality and synthesis.

Why is DINO v2 used in the proposed detection framework?

DINO v2 is used because its self-distillation enables strong feature extraction that captures both global artistic structures and fine-grained visual anomalies.

How do interpretability and statistical validation support the framework?

Interpretability methods such as Grad-CAM (and LIME mentioned) help explain model decisions, while statistical validation using log of p-values confirms dependable and transparent predictions.

Sep 30, 2026

Upload by Jacob