Loading...
Loading...
Contributors:
Dataset URI:
10.5281/zenodo.23538937This record contains the code, frozen protocols, trained models and evaluation results supporting the paper “Isolating the Effect of Training-Data Format Bias on Parameter-Efficient Fine-Tuning for AI-Generated Image Detection” (submitted to ICIMCIS 2026). The study tests whether removing encoding bias from training data changes the comparison between linear probing (LP) and low-rank adaptation (LoRA) for AI-generated image detection. In a two-by-two design, DINOv2 ViT-L/14 and, as an exploratory replication, CLIP ViT-L/14 are adapted by LP or LoRA on two versions of the same GenImage Stable Diffusion v1.4 images. In one version, real images are JPEG and synthetic images are PNG. In the other, synthetic images are JPEG-encoded with the quantization tables of paired real images. Models are evaluated on Synthbuster with RAISE-1k photographs, and on the WildRF test split. Included: Kaggle notebooks for dataset construction, training and evaluation; identifier lists with SHA-256 hashes for rebuilding both training versions; frozen protocol files; trained LoRA adapters and linear heads for every run; and per-image predictions, metrics and statistical analyses. The images themselves are not redistributed, because they belong to third-party datasets with their own terms of use (GenImage, Synthbuster, RAISE-1k, WildRF). See README.md for sources and reproduction steps.
Keyword(s):
AI-generated image detection; CLIP; DINOv2; encoding bias; JPEG compression.
MeSH Term(s):
Grants and Funding:
No Grants or Funding Data Available
Associated Publication(s):
No Associated Publication Data Available
Supplementary Data:
No Supplementary Data Available
Related Dataset(s):
No Related Dataset Data Available
The National Library of Medicine (NLM) provides information about biomedical datasets. Inclusion on an NLM Web page does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: NLM Web Policies (nih.gov)