Information Hiding by Machine Learning: A Method of Key Generation for Information Extracting Using Neural Network

Introduction

In this paper, we propose a method of key generation scheme for static visual information hiding by using machine learning technology, neural network as its exemplary approach for machine learning method.

The proposed method is to hide bit patterns as training data set for machine learning and the generated classifier acts as an information extraction key. These hidden patterns are prepared in advance to embedding procedure for the purpose of fingerprint or watermark signal of the target content. Machine learning is processed so that the generated classifier herein is neural network, to output the presented hidden patterns when feature values are presented to the classifier. This feature values are extracted using feature extraction key which is prepared before the embedding process. Feature extraction key contains information such as feature coordinates and regions of frequency coefficients. Also, the generated classifier in the machine learning process is saved as information extraction key for extraction process. Information extraction procedure is properly processed only by having two proper keys of feature extraction key and information hiding key.

Our method consists of three layers for digital watermarking as an application of our information hiding and information retrieval method: The first layer prepares feature extraction key, which is saved by feature extraction attributes including feature coordinates and regions of frequency coefficient of visual information; The second layer prepares watermark bit patterns to be related with the target content, in advance to the third embedding procedure, as watermark values of the visual information; And, the third layer generates watermark extraction key which is the classifier generated by the machine learning process. The discussed watermark extraction key and feature extraction key identify the related or associated hidden pattern which is the watermark values for proper digital watermarking procedure as shown in Figures 1 and 2.

Figure 1.

Key generation scheme in the proposed method

Figure 2.

Watermark extraction scheme in the proposed method

The proposed method is to contribute to secure visual information hiding without losing any detailed data of visual objects. The proposed method has used neural network for its training approach not limited but open in its applications to other machine learning approaches including fuzzy, Bayesian network and others. In this paper, the target content is a static visual data which are constructed with discrete data set and we have demonstrated the feasibility of solving this problem by using neural network model. We would enhance our method by using those other approaches, such as fuzzy for dynamic visual data like video stream data and Bayesian network for continuous data structures. This paper is different from the previous work by Ando and Takefuji (2003) in terms of embedding size where this paper does not embed any information to the target content and also implies that the machine learning algorithm is not limited only to the neural network model as proposed in our previous work (Naoe & Takefuji, 2008).