1. Digital Watermarking for Telltale Tamper Proofing and Authentication Deepa Kundur, Dimitrios Hatzinakos Presentation by Kin-chung Wong

2. Robust Watermark Able to recover embedded message from media even when attacked Application: –Authenticates content source –Digital rights management (DRM) –Embed secret message (steganography)

3. Fragile Watermark Embedded in such a way that, when the media is modified, the watermark tells something about the nature and strength of modification Application: tamper-proofing for court evidence, certificates, financial documents

4. Requirements 1.Indicate whether distortion exists. 2.Indicate relative magnitude of distortion 3.Characterize type of distortion, especially distinguishing between compression and intentional tampering 4.Validate and authenticate without generating metadata (non-image data sent along with an image file) … (*) (*) Uninformed users need not pass along metadata with the image. However, decoder still needs additional information from encoder, including original watermark bit sequence.

5. Requirements in details The watermark can be extracted from watermarked media (z) without explicit knowledge of host (f). Difference between w and provide information on signal modification, in terms of nature and strength of distortion User decides whether to accept content as authentic, based on the above information

6. Choice of embedding domain Discrete wavelet transform For N-bit watermark, select N coefficients in the wavelet domain, and embed one bit per coefficient using quantization index modulation Advantage: tampering is detected with both spatial and frequency localization by observing where errors occur most

7. Encoder Host DWT: Watermark: Coefficient selection key: Watermark quantization parameter (*): (*) different from, but related to coefficient rounding of file format

8. Decoder Extraction: get at the coefficient indicated by ckey(i) Tamper assessment function: (TAF) = (Number of bits flipped) / (Total number of bits)

9. Embedding one bit in a coefficient Quantized “bin” function Original wavelet coef.

10. Noise analysis Mild distortion –Small additive Gaussian noise – small, we can take approximations Severe distortion –Extracted watermark bits become unpredictable –Heavy filtering, least-significant-bit truncation, image region substitution, geometrical distortion –No correlation between w(i) and

11. Mild distortion Red: Changed but not detected Green: Changed and detected

12. Probability of not detecting a changed bit

13. Implementation: an example based on Haar wavelet Choosing based on coefficient rounding

14. Incorporating host dependence in watermark message Generate a quantization key qkey(i) from some image characteristics Perform XOR with qkey(i) and w(i) to get w*(i) to be embedded

15. Simulation: with mean filter M: mean filter lengths l: TAF detected at l-th DWT level (detail  … … …  approximate) TAF: tamper assessment function / fraction of watermark bits changed

16. Simulation: with JPEG compression

17. Simulation: image tampering (original image)

18. Simulation: image tampering (tampered image)

19. Bit error in watermark detected (at lower DWT level)

20. Bit error in watermark detected (at higher DWT level)

21. Discussions