Claudia Schremmer / University of Mannheim / Germany Digital Image Watermarking Claudia Schremmer University of Mannheim / Germany September 2001 Teleseminar on Electronic Commerce Introductory Week: in Nice

Claudia Schremmer / University of Mannheim / Germany page 2 Overview 1. Introduction 2. Digital Watermarking 3. Attacks 4. Algorithm of Zhao/Koch 5. Conclusion Overview

Claudia Schremmer / University of Mannheim / Germany page 3 Prelude Page from a prayer book showing the baking of unleavened bread (15th century) from the Illuminated Haggadah Exhibit.

Claudia Schremmer / University of Mannheim / Germany page 4 1. Introduction In former times... Traditional content media yield degraded content when copied: –paper documents, –analogue recordings, –celluloid film, –canvas paintings, and –marble sculptures Or they require expensive and specialized equipment to produce high-quality copies. Thus, the technical burden on traditional content creators for protecting their material has been minimal. 1. Introduction

Claudia Schremmer / University of Mannheim / Germany page 5 Nowadays... High quality (even: identical) copies are easy to produce. There is apparently no limit to the value that can be added by creating and providing access to digital content. Creators and owners of (digital) content need ways to protect their property rights in the increasingly digital world. Distributors and recipients of content need ways to fully utilize the rich potential of their digital assets. 1. Introduction

Claudia Schremmer / University of Mannheim / Germany page 6 Example Is this photo original? Is it fake / has it been modified? Who owns the right? Are we allowed to use this photo in our seminar? 1. Introduction Robbie Williams

Claudia Schremmer / University of Mannheim / Germany page 7 Definition 1. Introduction

Claudia Schremmer / University of Mannheim / Germany page 8 2. Digital Watermarking Digital watermarks are fixed to the underlying data. In contrast to the traditional way of watermarking paper or money, digital watermarks rely on steganography: A digital watermark is included into a digital medium the way that the user does not perceive it. Thus, the communication is secret. Two ways of secret communication: –The copyright information is transferred via a secret key while the algorithm is known. Thus, only the user can access the information. –No secret key exists, but the algorithm is kept secret. 2. Digital Watermarking

Claudia Schremmer / University of Mannheim / Germany page 9 Practical Approach Substitutional steganography: A noisy component of the digital message is substituted by an encrypted secret message. Constructive steganography: The secret message is not included by means of replacement, but by means of reproduction of noisy signals, based on the model of the original noise. Existing watermarking algorithms are based on the substitutional steganography. 2. Digital Watermarking

Claudia Schremmer / University of Mannheim / Germany page 10 Types of watermarks Subdivision of unperceptible (steganographic) watermarks: –Unperceptible robust: Insertion of hidden information, copyright information, authentification, and meta information. Based on a modification of the pixel (for an image), example: modification of an amplitude in the blue channel. Or based on a modification in the frequency domain, example: after DCT or wavelet transformation. –Unperceptible fragile: Aim to proof the integrity of a document. They are destroyed on purpose in order to detect manipulation. Based on content related algorithms like the digital signature. 2. Digital Watermarking

Claudia Schremmer / University of Mannheim / Germany page Attacks: unintentional There are a number of unintentional and intentional attacks: Unintentional attacks: –Image: compression, transcoding, printing/scanning, filtering, noise, geometric transforms, cropping, compositing/mosaicing,... –Video: AD/DA conversion, compression, transcoding, text/logo insertion, geometric transformations, jitter, cropping, Attacks

Claudia Schremmer / University of Mannheim / Germany page 12 Attacks: intentional Intentional attacks: –watermark removal/interference: denoising, compression, quantization, remodulation, blurring, averaging,... –Desynchronization (detector disabling): cropping, affine and projective transforms, jittering, mosaicing, collage,... –Cryptographic: key determination (brute force), Oracle attack (i.e., generate unmarked data by trial and error) –Protocol: copy attack, printing/rescanning,...  Watermark research must include work on attacks! 3. Attacks

Claudia Schremmer / University of Mannheim / Germany page 13 Example (I) 3. Attacks original wavelet compr.94%crumple & scan warpmosaic

Claudia Schremmer / University of Mannheim / Germany page 14 Example (II) 3. Attacks composition of wavelet compressed house and warped bear

Claudia Schremmer / University of Mannheim / Germany page 15 Constraints Human visual perception shall not be disturbed by the watermarks. The capacity for watermarks differs according to the number of bits allocated. Robustness regards the performance towards intentional and unintentional attacks. 3. Attacks visibility robustnesscapacity

Claudia Schremmer / University of Mannheim / Germany page Algorithm of Zhao/Koch Jian Zhao and Eckhard Koch: Embedding robust labels into images for copyright protection. In: Proc. of the International Congress on Intellectual Property Rights for Specialized Information, Knowledge and New Technologies. Vienna, Austria, August This is the initial paper which initialized a whole new research area. 4. Algorithm of Zhao/Koch

Claudia Schremmer / University of Mannheim / Germany page 17 Outline of the Zhao/Koch Algorithm The paper of Zhao/Koch describes a pool of novel steganographic methods for secure and robust insertion of watermarks into digital images. The included watermark is neither detectable, nor movable/relocatable, nor changeable. Furthermore, it survives attacks which do not strongly the image quality, such as lossless compression, low pass filtering (i.e., smoothing) and transcoding. 4. Algorithm of Zhao/Koch

Claudia Schremmer / University of Mannheim / Germany page 18 Fundamentals (I) System for Copyright Protection: SysCoP The system supports gray value, color, and binary images. 4. Algorithm of Zhao/Koch Labelled Image (y’) Position Sequence Generator Ts(y, U k ) Label Embeddin g System Position Sequence (PS) Image Data (y) User Key (U k ) Label Code (C) write a watermark Two steps to write a watermark: –Generation of a pseudo-random sequence of positions in order to select DCT-encoded blocks in which the code is embedded. Here, this step is accomplished by the function –Embedding of the code into the block selected by the sequence of positions.

Claudia Schremmer / University of Mannheim / Germany page 19 Fundamentals (II) The reading process reverses the writing process: –Reading the code from the block selected by the sequence of positions. –Retrieve the label inserted. 4. Algorithm of Zhao/Koch Position Sequence Generator Ts(y, U k ) Label Retrieval System Position Sequence (PS) Image Data (y) User Key (U k ) Embedded Code (C) read a watermark

Claudia Schremmer / University of Mannheim / Germany page 20 Remarks In the following: consideration of gray value images only. A block is of size 8x8 pixels. It can be contiguous or distributed. –A contiguous block is a square part of an image. –A distributed block is a random agglomeration of 64 pixels of the image. It is difficult to determine an optimal function. Therefore, only the width and the height of an image are actually used to generate the block positions. 4. Algorithm of Zhao/Koch

Claudia Schremmer / University of Mannheim / Germany page 21 Framework(write) Algorithm: Framework(write). –(1) If i  n, return. –(2) Randomly select a block b, using the position sequence generation function Ts(U k, y). –(3) If b exists already in B, go to (2), otherwise add b to B. –(4) Call check_write(b, c i ) to check weather b is a valid block: if this function returns False, go to (2). –(5) Call write(b, c i ) to embed a bit c i to the block b. –(6) Increment i, go to (1). 4. Algorithm of Zhao/Koch –The code C to be embedded is represented by its binary bit sequence: –i marks the current bit in the sequence –B is the set of blocks of which a block is randomly selected. –The initialization is: i=0, B={}.

Claudia Schremmer / University of Mannheim / Germany page 22 Framework(read) Algorithm: Framework(read). –(1) If i  n, return. –(2) Randomly select a distributed or a contiguous 8  8 block b, using the position sequence generation function Ts(U k,y). –(3) If b exists already in B, then go to (2), otherwise add b to B. –(4) Call check_read(b, c i ) to check weather b is a valid block: if this function returns False, go to (2). –(5) Call read(b) to retrieve a bit from the block b. –(6) Increment i, and go to (1). 4. Algorithm of Zhao/Koch

Claudia Schremmer / University of Mannheim / Germany page 23 JPEG-based Code Insertion In the framework, the following algorithms are still undefined: –check_write(b, c i ) –write(b, c i ) –check_read(b, c i ) –read(b, c i ) They depend on the coding format of the image. We restrict our future considerations to JPEG-encoded images: 4. Algorithm of Zhao/Koch Image pre- processor Pixel Block Image encoder DCT QuantizationEntropy encoding: run length Huffman arithmetic

Claudia Schremmer / University of Mannheim / Germany page 24 Visual Perception General idea: Insert watermarks in positions where the human visual system does not or nearly not perceive a modification of the coefficients. Here: Insertion in the smooth background areas, i.e., in the white areas of the right picture. The coefficients concerned are of the middle frequencies in the DCT-transformed domain. 4. Algorithm of Zhao/Koch

Claudia Schremmer / University of Mannheim / Germany page 25 Possible Positions Statistic and empirical research has defined possible positions to include a code into a block. 4. Algorithm of Zhao/Koch

Claudia Schremmer / University of Mannheim / Germany page 26 Demonstration 4. Algorithm of Zhao/Koch

Claudia Schremmer / University of Mannheim / Germany page Conclusion We have not... –... detailed any precise algorithm. –... discussed other media than images. For audio, this would e.g. mean entering the field of research on human audible properties. –... discussed the various attacking algorithms like e.g. STIRMARK. We have... –... given an introduction and an overview on the purpose of digital watermarking. –... outlined the challenge of creating a robust algorithms to unintentional and intentional attacks. –... restricted our considerations on images. 5. Conclusion

Claudia Schremmer / University of Mannheim / Germany page 28 Information... Claudia Schremmer University of Mannheim, Germany Department Praktische Informatik IV