1. Concepts of Multimedia Processing and Transmission IT 481, Lecture #9 Dennis McCaughey, Ph.D. 2 April, 2007

2. 04/02/2007 IT 481, Spring 2007 2Project Project will be recovering a student specific watermark from the bird image. MATLAB code will be provided you only have to execute it. This is not a programming project. e-mail me at dennisgmc1@verizon.netdennisgmc1@verizon.net – I will reply with the code. Criteria: –Recover the watermark –Assess the visibility and robustness of the watermarking method

3. 04/02/2007 IT 481, Spring 2007 3 Homework #4 E-mail “dennisgmc1@verizon.net” requesting MATLAB code and “bird” picture. I will reply with the code. Load these on a CD and bring to 4/9 class Review the following instructions –Insert the CD –Open Matlab –Under “file” open D\:watermarkread.m –Under “debug” click “run” –When prompted, set the search path to the top of the Matlab search order, click “ok” –When prompted, enter your student ID # Project –Answer the following questions What is the embedded message? How many errors are reported? The number of errors may not equal the misspelled words, Why? –E-mail your answers to me by 4/13

4. 04/02/2007 IT 481, Spring 2007 4Watermarking Watermarking is a secret code described by a digital signal carrying information about the copyright property of the product. The watermark is embedded in the digital data in such a way that it is not visually perceptible. The copyright owner should be the only person who can show the existence of his own watermark and to prove then origin of the product.

5. 04/02/2007 IT 481, Spring 2007 5 Watermark Requirements Alterations introduced into the image or audio should be perceptually invisible. A water mark must be undetectable and not removable by an attacker. A sufficient number of watermarks in the same image or audio, detectable by their own key, can be produced. The detection of the watermark should not require the original image or audio. A watermark should be robust against attacks which preserve the desired quality of the image or audio.

6. 04/02/2007 IT 481, Spring 2007 6 Main Features of Watermarking Perceptual Invisibility Trustworthy Detection Associated Key Automated Detection/Search Statistical Invisibility Multiple Watermarks Robustness

7. 04/02/2007 IT 481, Spring 2007 7 Perceptual Invisibility Watermark should not degrade the perceived image/audio quality Differences may become apparent when the original and watermarked versions are directly compared

8. 04/02/2007 IT 481, Spring 2007 8 Trustworthy Detection Watermarks should constitute a sufficient and trustworthy part of ownership. False alarms should be extremely rare. Watermarks signatures/signals should be complex. An enormous set of watermarks prevents recovery by trial-and-error methods.

9. 04/02/2007 IT 481, Spring 2007 9 Associated key Watermarks should be associated with an identifiable number called the watermark key. Key used to cast, detect and remove the watermark. The key should be private and should exclusively characterize the legal owner. Any signal removed from the image/audio is assumed to be valid only if it can be associated with the key via a well established algorithm

10. 04/02/2007 IT 481, Spring 2007 10 Automated Detection/Search Watermark should combine with a search algorithm.

11. 04/02/2007 IT 481, Spring 2007 11 Statistical Invisibility Watermark should not be recoverable using statistical methods. The possession of a great number of watermarked images, embedded with the same key should not enable the recovery of the watermark through statistical methods. –Watermarks should be image/audio independent.

12. 04/02/2007 IT 481, Spring 2007 12 Multiple Watermarks Multiple watermarks assist in the case where someone illicitly watermarks and already watermarked image/audio. Convenient in transferring copyrighted material.

13. 04/02/2007 IT 481, Spring 2007 13Robustness A watermark should survive some modifications to the data. Common manipulations to image/video –Data Compression –Filtering –Color, quantization, brightness modifications, geometric distortions, etc –Other trans-coding operations.

14. 04/02/2007 IT 481, Spring 2007 14 Application Domains A1: Carrying value-added metadata –Additional information such as hyperlinks, content based indexing –Malicious and non-malicious attacks –Survive MPEG encoding A2: Copy protection and conditional access –Control Intellectual Property Management and Protection –View and copy options –Every compliant decoder must be able to trigger protection or royalty collection mechanisms at the time of decoding –Unauthorized individuals should not be able to defeat the watermarks by any means A3: Ownership assertion, recipient tracking –Establish ownership and determine origin of unauthorized duplication. –Prosecution of copyright infringement

15. 04/02/2007 IT 481, Spring 2007 15 Application Domains Cont’d A4: Authentication and verification –Allows fragile watermarks; if contents modified watermarks should disappear. –Helps in identifying areas that wer modified A5: Broadcast monitoring –Monitor where and when the contents are played –Advertisements. Here heavy content degradation is less of an issue. –Watermark removal, invalidation and forgery can be significant concern –Counterfeiting should be intractable for the system to be effective A6: Secret communication or steganography –Data hiding may require higher capacity watermarks than other applications –Secrecy may be the overriding concern in some applications

16. 04/02/2007 IT 481, Spring 2007 16Attacks AT1: Basic attacks –Lossy compression, frame dropping & temporal rescaling AT2: Simple attacks –Blurring, median filtering, noise addition gamma correction and sharpening AT3: Normal attacks –Translation, cropping and scaling AT4: Enhanced attacks –Aspect ratio change & random geometric perturbations (Stirmark) AT5: Advanced Attacks –Delete/insert watermarks, single document watermark estimation attacks & multiple-document statistical attacks

17. 04/02/2007 IT 481, Spring 2007 17 Robustness, Resilience & Detection Applications Domain Unintentional Attacks Intentional Attacks Every Decoder High Capacity Applications Example AT1AT2AT3AT4AT5 A1Yes MaybeNo Yes Value-added metadata A2Yes NoCopy Protection A3Yes No Ownership/fingerprint A4YesNo SomeYesNoAuthentication A5Yes NONoYes Broadcasting A6Yes Maybe YesNoYes Secret Communication

18. 04/02/2007 IT 481, Spring 2007 18 Human Perception Watermarking schemes take advantage of the fact that the human audio and visual systems are imperfect detectors. Audio & visual signals must have a minimum intensity or contrast before they are perceptible. These minima are spatially, temporally and frequency dependent. These dependencies are either implicitly or explicitly exploited

19. 04/02/2007 IT 481, Spring 2007 19 Transform Domain Considerations The human eye is more sensitive to noise in the lower frequency range than in the higher frequency counterparts However, energy in most images is concentrated in the lower frequency range. Quantization used in DCT based compression reflects the HVS which is less sensitive in the higher frequencies A trade is required to balance watermark invisibility and survivability resulting in the use of the mid-frequency terms.

20. 04/02/2007 IT 481, Spring 2007 20 Transform Domain Considerations An alteration of a transform coefficient is spread across the entire spatial block A one dimensional example:

21. 04/02/2007 IT 481, Spring 2007 21 Data Embedding Algorithm

22. 04/02/2007 IT 481, Spring 2007 22 Embedded Data Examples Multilingual soundtracks within a motion picture Copyright data Distribution permissions Data used for accounting and billing and royalties Etc.

23. 04/02/2007 IT 481, Spring 2007 23 Watermarking Techniques Non-Blind: Watermark recovery requires the original Blind: Watermark recovery does not require the original Spatial domain or transform domain embedding Spatial domain: –LSB, color pallet, geometric Transform Domain: –FFT, DCT, Wavelet

24. 04/02/2007 IT 481, Spring 2007 24 An Algorithm Proposed by Busch, Funk and Wolthusen

25. 04/02/2007 IT 481, Spring 2007 25 Algorithm Considerations Watermark embedding position determined through a pseudo-random number generator that determines the order of block processing and the coefficient to be modified –Embed all available blocks Key may be public or secret leading to a public or secret watermark Redundantly embed the watermark to achieve survivability to MPEG-2 compression

26. 04/02/2007 IT 481, Spring 2007 26 Selected Block Embedding 1. Block is transformed using the DCT 2. Perform edge detection 3. Select pair of DCT coefficients from the sub-band used using a Pseudo random permutation 4. Quantize the selected coefficients using the MPEG-2 algorithm 5. Determine if the coefficient pair is suitable for embedding Avoid “Edge” and “Plain” blocks 6. Enforce a differential relationship between the coefficients in the pair to encode a “1” or a “0” 7. Inverse quantize the modified coefficients

27. 04/02/2007 IT 481, Spring 2007 27 Busch Watermarking DCT Coefficients If the absolute value of one of the highlighted coefficients is greater than 39, the block is classified as an “edge” and not used. If the quantized value of one of the coefficients in the selected band is zero the block is classified as a “plain” block

28. 04/02/2007 IT 481, Spring 2007 28Visibility “Edge” blocks, if modified, are highly visible in video and are to be avoided “Plain” blocks are not so sensitive, so they can be used if care is exercised –In one of the bands randomly select a pair of coefficients and randomly select one to be the “first” (DCT1) and another to the “second” (DCT2) –To encode a “one” set DCT1 = (ABS(DCT1)+ ABS(DCT2))/2 + K, Preserving the sign(DCT1) –To encode a “zero” set DCT1 = (ABS(DCT1)+ ABS(DCT2))/2 - K, Preserving the sign(DCT1) –K is elected as a compromise between visibility and robustness to MPEG-2 compression

29. 04/02/2007 IT 481, Spring 2007 29Reference C. Busch, W. Funk, and S. Wolthusen: “Digital Watermarking: From Concepts to Real-Time Video Applications”; IEEE Computer Graphics and Applications, 1999

30. Matlab Overview