[Volume 1, Issue 1] 8. Forensics towards image super-resolution generated by deep networks
2021-04-07 Haichao Yao, Rongrong Ni*, and Yao Zhao* 

Abstract:

With the help of deep networks, image super-resolution (SR) has developed rapidly in recent years. However, a low-resolution image can be restored to a realistic high-resolution one by means of the SR methods, that could lead to security and copyright issues. To the best of our knowledge, no specific forensic technique has yet been proposed to identify whether a high-resolution image is SR image or not.

In this paper, the forensics of image SR generated by deep networks is studied.

We observe that the SR images show re-executed stability. When the SR image is executed by SR methods once again, the differences between the double SR and the single SR image are relatively small in the frequency domain, but the differences between the single SR image and the corresponding original image are relatively large.

According to this observation, a 40-D feature based on the Histogram of Power Spectral Density Residuals (H-PSDR) is proposed for SR detection.

In order to achieve uniform training and test via the proposed feature, we design a general detection framework which make full use of the known SR models for training and test.

Compared with the typical feature based forensic methods and deep classification network, our proposed method has strong generalization capability and is able to overcome the mismatch of different SR models in training and test. The experimental results show that our proposed method performs well in different situations.

Keywords:

Digital forensics, Image super-resolution, H-PSDR feature, Mismatch training.

Authors and contacts:

Haichao Yao, Beijing Key Laboratory of Advanced Information Science and Network Technology, Beijing 100044, China;

Institute of Information Science, Beijing Jiaotong University, Beijing 100044, China, 17112062@bjtu.edu.cn

Rongrong Ni*, Beijing Key Laboratory of Advanced Information Science and Network Technology, Beijing 100044, China;

Institute of Information Science, Beijing Jiaotong University, Beijing 100044, China, rrni@bjtu.edu.cn

Yao Zhao*, Beijing Key Laboratory of Advanced Information Science and Network Technology, Beijing 100044, China;

Institute of Information Science, Beijing Jiaotong University, Beijing 100044, China, yzhao@bjtu.edu.cn

Citation

Haichao Yao, Rongrong Ni, and Yao Zhao, “Forensics towards image super-resolution generated by deep networks”, Journal of Computer Security and Data Forensics, Vol. 1, No. 1, pp. 109~126, March 2021.