Watermarking based Image Authentication using Feature Amplification Shuiming Ye 1, 2 , Ee-Chien Chang 1 and Qibin Sun 2 1 School of Computing, National University of Singapore, Singapore 2 Institute for Infocomm Research, A*STAR, Singapore Abstract In a typical content and watermarking based image authentication approach, a feature is extracted from the given image, and then embedded back into the image using a watermarking method. Since the entropy of the feature might be higher than the capacity of the wa- termarking scheme, or the feature is represented in a continuous domain, it has to be further quantized before embedding. The lost of information during quantization potentially degrades the overall performance of the authentication scheme. This paper propose a simple but effective approach that avoids the feature quantization by additive feature: the feature is firstly added into the image before watermark embedding, and latterly sub- tracted from the watermarked image. In our experi- ments, the proposed approach obtains larger achiev- able robustness/sensitivity region and has a smaller fuzzy region of authenticity than the typical approach. 1. Introduction The pervasive distribution of digital images triggers an emergent need of authenticating images received by unreliable channels, such as public Internet and wireless networks. Many image authentication techniques have been proposed, which can be divided into two broad categories: labeling based approaches [1, 2, 3], in which authentication data are written in a separate file; and watermarking based approaches [4, 5], in which au- thentication data are embedded in the images. This paper will focus on the watermarking based methods. A typical approach of content and watermarking based authentication follows three steps: firstly, a fea- ture is extracted from the image; secondly, the feature is discredited and quantized; finally, the quantized feature is embedded as message into the image. During au- thenticity verification, the message is detected using the watermarking detector, and the feature is extracted from the watermarked image. A typical authenticity verifica- tion decision is based on the comparison between a preset threshold T and the distance of the extracted feature and the detected watermark. Assuming the feature representation is a good choice, the performance of the above scheme depends on the effectiveness of the feature quantization and watermarking code. In practice it is impossible to obtain a perfect code. As a result, there is often a fuzzy region separates the space of authentic images from the unau- thentic images [6]. In this paper we propose an alternative approach by avoiding the quantization step. In doing so, potentially we can reduce the size of the fuzzy region. In our ap- proach, the feature is not quantized, but added into the image before watermarking. The feature is then sub- tracted from the watermarked image to get the output of the embedding procedure. Experimental results show that compared with the typical approaches, our scheme can achieve a smaller fuzzy region of authenticity, and a better achievable robustness/sensitivity region. 2. Proposed image authentication using feature amplification Fig. 1 shows the structure of the proposed approach. It is a general scheme in the sense that different wa- termarking schemes and feature extractions can be em- ployed. The embedding and verifying procedures re- quire a predefined parameter, the amplifying factor k, and a predefined message m. Watermarked Image Feature Extraction Watermark Embedding + - k k Original Image m (a) Embedding Watermark Detecting c > T? Authentic? Yes/No m Feature Extraction + k Received Image c (b) Authenticity verification Fig. 1: Proposed feature aided authentication scheme 2.1. Embedding Given an image I, the embedding procedure outputs 0-7803-9332-5/05/$20.00 ©2005 IEEE