The Dresden Image Database for Benchmarking Digital Image Forensics

ABSTRACT

This article introduces and documents a novel image database specifically built for the purpose of development and benchmarking of camera-based digital forensic techniques. More than 14,000 images of various indoor and outdoor scenes have been acquired under controlled and thus widely comparable conditions from altogether 73 digital cameras. The cameras were drawn from only 25 different models to ensure that device-specific and model-specific characteristics can be disentangled and studied separately, as validated with results in this article. In addition, auxiliary images for the estimation of device-specific sensor noise pattern were collected for each camera. Another subset of images to study model-specific JPEG compression algorithms has been compiled for each model. The Dresden Image Database is freely available for scientific purposes. The database is intended to become a useful resource for researchers and forensic investigators. Using a standard database as a benchmark makes results more comparable and more reproducible, and it is more economical. It also avoids potential copyright and privacy risks of self-collected benchmark sets downloaded from photo-sharing sites on the Internet.

Keywords:

• digital image forensics
• benchmarking
• image source identification
• camera model identification
• forgery detection

ACKNOWLEDGEMENT

The authors gratefully acknowledge the staff members of the faculty of computer science, the staff members of the AVMZ (TU Dresden), and the following companies (in alphabetical order) for borrowing us their digital cameras: AgfaPhoto & Plawa, Casio, FujiFilm, Kodak, Nikon, Olympus, Pentax, Praktica, Ricoh, Rollei, Samsung and Sony. Special thanks belong to the administration of Dresden Palaces and Gardens and to the administration of the Sculpture Collection Dresden for granting photography permissions in their premises. We are indebted to our colleagues and friends who supported the creation of the image database actively during the coldest period of the year 2009.

The conference version of this article appeared in the Proceedings of the 25th Symposium On Applied Computing (ACM SAC 2010), Vol. 2, pages 1585–1591.

Notes

1. A commendable exception is the image splicing database by Hsu and Chang [13]. However, it is limited to the detection of very basic attempts of manipulations. It includes only a small set of images and it does not control for scene content.

2. To benchmark forgery detection algorithms, additional fake images are required. The proposed database can be a starting point to create such manipulated images [14], though manipulated images are not part of the database yet.

3. We employed the support vector machine libSVM with a radial-based kernel function to separate different camera models [15].