edltauSCI

Smartphone photos for source camera identification, collected for the paper "Evaluating Deep Learning and Traditional Approaches Used in Source Camera Identification" (Mansur Ozaman, Nazarbayev University), accepted at SIDe'26 and to appear in the Springer STEAM-H series. Paper: arXiv:2511.19180. Code: https://github.com/ozxmn/edltauSCI. Archived copy: doi:10.5281/zenodo.17640511.

Contents

68 photos, 17 per device.

File	Device	Format
Ip13promax.zip	iPhone 13 Pro Max	HEIC
Ip17pro.zip	iPhone 17 Pro	HEIC
Samsungs24.zip	Samsung Galaxy S24	JPEG
redminote10s.zip	Xiaomi Redmi Note 10S	JPEG

All photos were taken in everyday conditions with the default camera app settings, the main rear camera and 1x zoom, and copied to a computer over a cable, so no messaging app or cloud service recompressed them.

The two formats differ by device, so a classifier can partly learn the file format instead of the sensor. The paper measures this with same-format device pairs and with every image re-encoded by one JPEG encoder. The Samsung files are stored rotated and carry an EXIF orientation tag; the code rotates every image upright before training.

Results in the paper

Mean accuracy over the same 25 folds of repeated stratified 5-fold cross-validation:

Method	Accuracy
DCT/JPEG artifact features + SVM	0.947
ResNet-18 (ImageNet, fine-tuned)	0.623
PRNU-inspired residual features + SVM	0.473
Compact CNN from scratch	0.356 to 0.411

17 images per device is a small sample: the confidence intervals in the paper are wide, and the results describe these four phones, not phone models in general.

License

CC BY 4.0.

Citation

@misc{ozaman2025sci,
  title         = {Evaluating Deep Learning and Traditional Approaches Used in Source Camera Identification},
  author        = {Ozaman, Mansur},
  year          = {2025},
  eprint        = {2511.19180},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2511.19180}
}