Work

Artifact-Guided Logit Fusion of Fine-tuned CNN and ViT Branches for AI-Generated Image Detection: Implications for AI Literacy in Informatics Education

Year: 2026
Type: article
Abstract: Background: The rapid advancement of generative artificial intelligence has made it increasingly difficult to distinguish authentic images from synthetic visual content, raising concerns for digital and AI literacy and academic integrity in higher education. Informatics curricula increasingly need concrete, evaluated examples of how AI-generated image detectors are designed, tested, and limited. Purpose: This study proposes Artifact-Guided EfficientViT-V3 (AG EfficientViT V3), a hybrid detection framework integrating fine-tuned EfficientNetB0 and ViT-Tiny branches with an artifact-oriented branch through learnable logit-level fusion, and discusses its potential relevance as an instructional case study and a supporting tool for AI/digital literacy and academic-integrity practice in informatics education. Method: Instead of direct feature concatenation, the model preserves the decision behavior of separately fine-tuned branches while adding an artifact-sensitive correction signal at the decision level. Results: On the CIFAKE benchmark, AG EfficientViT V3 achieves 98.865% accuracy, 98.938% precision, 98.790% recall, 98.864% F1-score, and 0.999116 ROC-AUC, the strongest clean-test performance among all evaluated baselines and variants. Ablation results show that naive CNN-Transformer hybridization alone does not outperform the strongest single-branch baseline, whereas fine-tuned branch initialization combined with logit-level fusion does. Robustness evaluation shows strong performance under clean and JPEG-compressed conditions, while severe blur, resize degradation, and additive noise remain challenging, illustrating that detector accuracy is condition-dependent rather than universal. Conclusion/Implications: The proposed pipeline and evaluation protocol can be adapted as teaching material in deep learning and computer vision courses, while integration into digital-literacy or academic-integrity workflows is framed as a potential application requiring further validation rather than an established outcome.
Language: English
Open Access status: diamond
OpenAlex ID: W7204498854
FWCI:
Cites: 24
Cited by:
Related to: 10