Employing intelligence in the embedding and decoding stages of a robust watermarking system

Abstract

This letter presents a novel approach of incorporating intelligence in the encoding and decoding structures of a watermarking system. The employment of computational intelligence makes the watermarking system resistant against a series of attacks, which may occur during the storage or communication of the watermarked work. Keeping in view the Human Visual System, Genetic Programming is used to generate functions which select optimum strength and location of transform domain coefficients for watermark embedding. Support Vector Machines and Artificial Neural Networks are employed at the decoding side to learn about the distortions due to attacks and counteract them. Especially, the proposed system is quite effective for robust watermarking applications of small size images such as those displayed on portable devices and online catalogs.

Introduction

Digital watermarking has gained significant popularity within the research community in recent years due to its considerable advantages over the traditional data hiding techniques especially as regards copyright protection [1], [2] and authentication related applications [3]. In case of copyright protection based application, watermarking systems still face the challenge of being robust against real world attack scenarios, which consist of a series of legitimate/illegitimate distortions [4]. These distortions are somewhat conceivable based on application environment and the level of access the attacker has. A watermark is usually concealed in the cover work by keeping in view the response of human perception mechanism. For this reason, it is amplified in the regions where it is imperceptible to the naked eye and attenuated in areas where it could be perceptible. This is referred as perceptual shaping and incorporated by the help of perceptual models based on Human Visual System (HVS). Watson's Perceptual Model (WPM) [5], [6] is a pertinent example along many others proposed in literature [7], [8], [9], [10]. However, these perceptual models are based on visual characteristics alone and altogether lack the shaping ability in view of attacks. This is primarily due to the reason that attack information is not taken into account while designing these models. Therefore, it is judicious to propose new Perceptual Shaping Functions (PSFs) that are able to shape a watermark with respect to its robustness against a series of real world attacks, as well as maintaining its invisibility property. Our approach for developing such PSFs is by employing multi objective Genetic Programming (GP) [11] to generate expressions (PSFs) that make a delicate balance between robustness and imperceptibility at the watermark embedding stage. Through GP simulation, PSFs are evolved by stochastically searching for appropriate coefficient positions and suitable strength of alterations for watermark embedding in Discrete Cosine Transform (DCT) domain. Objective measures for watermark robustness and imperceptibility comprise the fitness function in order to drive the simulation towards an optimum/pareto-optimal region. Similarly, an intelligent watermark extraction strategy is needed at the decoding stage which may adapt itself in accordance to the widespread and dynamic applications of watermarking, where different types of attacks are expected at different instances. In this context, most of the existing watermark extraction strategies [12], [13], [14] do not consider the existence of attacks during the training phase and thus are not adaptive. In the same way, other approaches, such as threshold based decoding techniques [15], [16] are also not versatile towards attacks. These approaches neither considers the alterations that may incur to the features nor exploit the individual frequency bands; rather treat all the frequency bands collectively. Consequently, these approaches use a single feature for extraction of a message bit. Recently, Khan et al. [17] have proposed machine learning (ML) based decoding. However, their system does not employ intelligent embedding. In this letter we propose a novel scheme which utilizes ML techniques at the watermark embedding as well as decoding stages. For this purpose, firstly, we perform GP based intelligent embedding which selects appropriate frequency bands and the magnitude of allowable alterations for watermark embedding. Considering the available frequency bands in an 8 × 8 block DCT, this in turn also acts as a feature reduction strategy. Secondly, we exploit the learning capabilities of Support Vector Machine (SVM) and Artificial Neural Networks (ANN) to gain knowledge of the distortion that might have incurred varyingly on the different selected frequency bands due to the sequence of attacks. Lastly, to perform this efficiently and in an adaptive manner, we devise a new feature extraction strategy for watermark extraction in presence of attacks. Both of the above attributes of the proposed scheme make it extremely difficult for an adversary to gain the secret knowledge of the system by analyzing the embedding and/or decoding space. The paper is organized as follows. Section 2 describes the proposed watermarking system including details of the encoding and decoding phases. Our experimental results are discussed in Section 3 followed by conclusions in Section 4.