XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
A Robust Image Zero-watermarking using
Convolutional Neural Networks
Atoany Fierro-Radilla
ESIME Culhuacan
Instituto Politecnico Nacional
Mexico City, Mexico
afierror@hotmail.com
Laura Cleofas-Sanchez
ESIME Culhuacan
Instituto Politecnico Nacional
Mexico City, Mexico
laura18cs@hotmail.com
Mariko Nakano-Miyatake
ESIME Culhuacan
Instituto Politecnico Nacional
Mexico City, Mexico
mnakano@ipn.mx
ORCID:0000-0003-1346-7825
Hector Perez-Meana
ESIME Culhuacan
Instituto Politecnico Nacional
Mexico City, Mexico
hmperezm@ipn.mx
ORCID: 0000-0002-7786-2050
Manuel Cedillo-Hernandez
ESIME Culhuacan
Instituto Politecnico Nacional
Mexico City, Mexico
mcedillohdz@hotmail.com
ORCID: 0000-0002-9149-9841
Abstract— In the image zero-watermarking techniques, a
watermark sequence is not physically embedded into the host
image but has a logical linkage with the host image. This property
of zero-watermarking is desirable for some kinds of images in
which a minimum distortion may cause serious detection or
diagnostic errors, such as medical images and remote sensing
images. In this paper, we propose a robust zero-watermarking
algorithm based on the Convolutional Neural Networks (CNN)
and deep learning algorithm, in which robust inherent features of
image is generated by the CNN, and it is combined with the
owner’s watermark sequence using XOR operation. The
experimental results show the watermark robustness against
several attacks and common image processing.
Keywords—Zero-watermarking, Deep Learning, Convolutional
Neural Networks, Robust Features
I. INTRODUCTION
During last two decades, several types of image
watermarking algorithms have been proposed for different
purposes, such as the copyright protection [1, 2], owner
identification [3, 4] and content authentication [5, 6]. Almost all
of them, a watermark sequence is physically embedded into the
host image, causing some distortion in it. In the invisible
watermarking, the distortion caused by the embedded
watermark is almost imperceptible by the Human Visual System
(HVS). However, in some kind of images, such as medical
images and remote sensing images, a small distortion may
causes serious diagnostic or detection errors.
Until now two kinds of distortion-free watermarking
techniques have been proposed in the literature. The first one is
reversible watermarking, in which once watermark sequence is
detected or extracted from the watermarked image, the original
undistorted host image can be recovered [7, 8]. Generally, the
reversible watermarking techniques cannot provide sufficient
robustness of watermark sequence to common signal
processing, such as JPEG compression and noise contamination.
The second distortion-free watermarking technique is zero-
watermarking, in which the watermark sequence is not
physically embedded into the host image, but logically linked
with the host image, keeping the host image intact.
In the zero-watermarking, instead of embedding a
watermark sequence or watermark pattern, some inherent
features are extracted from the host image. These inherent
features are linked with an owner’s watermark sequence to
generate a master share, which is stored in a secure manner [9-
13]. The protected images by the zero-watermarking are
transmitted through any insecure public communication channel
and the owner can verify his ownership using the master share
and the inherent features extracted from the image under
analysis. In the zero-watermarking technique, the extraction of
robust inherent features of the host image is the most important
issue for their desirable performance.
In [9, 10], using Singular Value Decomposition (SVD),
some largest singular values are obtained from the
approximation sub-band (LL sub-band) after the 2D Discrete
Wavelet Transform (DWT) decomposition is applied to the host
image. The singular values are linked with the watermark
sequence using XOR operation and stored as master share. In
[11], first the host image is normalized using Hu’s image
normalization technique [14] and applied the SVD in the
Contourlet domain of the normalized host image to extract
robust features. In [12], the image normalization technique is
used, and the Bessel-Fourier moment is obtained from the
normalized image, which is used to generate a master share. In
[13], some robust Quaternion Exponent moments are extracted
from the host image to generate master share.
Recently, the Convolutional Neural Networks (CNN)
together with deep learning algorithms are used to solve several
computer vision problems, in which the CNN is trained to
extract the useful features of image for a desired task. Unlike the
conventional methods, in which the hand-crafted features are
firstly extracted and then some classifiers are trained using pre-
obtained hand-crafted features, in the CNN-based approach, the
CB2014-0237925, Mexican National Council of Science and Technology
