Abstract
It is well known that the JPEG compression pipeline leaves residual traces in the compressed images that are useful for forensic investigations. Through the analysis of such insights the history of a digital image can be reconstructed by means of First Quantization Estimations (FQE), often employed for the camera model identification (CMI) task. In this paper, a novel FQE technique for JPEG double compressed images is proposed which employs a mixed approach based on Machine Learning and statistical analysis. The proposed method was designed to work in the aligned case (i.e., $8 \times 8$ JPEG grid is not misaligned among the various compressions) and demonstrated to be able to work effectively in different challenging scenarios (small input patches, custom quantization tables) without strong a-priori assumptions, surpassing state-of-the-art solutions. Finally, an in-depth analysis on the impact of image input sizes, dataset image resolutions, custom quantization tables and different Discrete Cosine Transform (DCT) implementations was carried out.
Highlights
The everyday number of digital images acquired, stored or shared is constantly growing due to the widespread of social networks
To other kind of data, it is possible to give a life-cycle to digital images, consisting on the following steps: acquisition by means of a digital device, editing and uploading to Instant Messaging platforms or Social Networks
The aforementioned steps, produce a JPEG double compression [1] on the image data, which means that a lot of the original information is definitively lost
Summary
The everyday number of digital images acquired, stored or shared is constantly growing due to the widespread of social networks. The proposed method was designed to cope with the typical limits affecting both statistical FQE approaches (e.g., low accuracy with small patches, strong a-priori assumptions about involved quantization matrices) and machine learning based methods (e.g., overfitting with respect to input patch content and quantization tables employed in the training phase). It works without strong a-priori assumptions on the quantization matrices employed in the compression process of the analysed images.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
