1. Damageless Information Hiding Technique using Neural Network Keio University Graduate School of Media and Governance Kensuke Naoe

2. Abstract An information hiding technique without embedding any data to target content  Pattern recognition model Neural network as classifier (extraction key)  Advantage and disadvantage

3. Outline Background Motivation Current Problem Proposed Method Experiment results Future Work Coclusion

4. Background Emergence of the Internet  Contents are widely distributed Information hiding provides reliability  Digital watermarking for Digital Rights Management  Steganography for covert channel

5. Motivation and current problem Use one information hiding algorithm with another to strengthen the security of the content  Digital watermarking  Steganography  FIngerprinting There are many great information hiding algorithm but have difficulties to collaborate  possibility of obstructing previously embedded data  Applying another information hiding algorithm might result in recalculation of fingerprint for the content

6. Research Objective To hide or to relate certain information without embedding any information to the target content Ability to collaborate with another information hiding algorithm to strengthen the security

7. Proposed Method Approach  Embed model to pattern recognition model Neural network as classifier (extraction key)  Only proper extraction key will lead to proper hidden signal

8. Why use neural network? Has abilities of  Tolerance to noise  Error correction and complementation  Additional learning characteristic Multi-layered Perceptron Model  Backpropagation Learning (Supervised Learning)

9. Proposed Method (Embedding) 1.Frequency Transformation of content Hidden signal asteacher signal 2.Selection of feature subblock 3.Use feature values as input value for neural network 4. Generation of classifier (extraction key) Coordinate of feature subblocks (extraction key)

10. Proposed Method (Extraction) 1.Frequency Transformation of content Hidden signal asoutput signal 2.Selection of feature subblock 3.Use feature values as input value for neural network 4. Applying the classifier (encryption key) Coordinate of feature subblocks (encryption key)

11. What is neural network? neuron （ nervous cell ）  It only has a function of receiving a signal and dispatching signal to connected neuron  When organically connected, it has ability to process a complicated task A network built with these neurons are called neural network  Multi layered perceptron model Often used for non-linear pattern classifier

12. Calculation of network Input value of neuron  Sum product of network weight and output values from previous layer j xjxj yjyj y1y1 yiyi yNyN w 1j w ij wNjwNj

13. Generating classifier (extraction key) 1.Frequency Transformation of content Hidden signal asteacher signal 2.Selection of feature subblock 3.Use feature values as input value for neural network 4. Generation of classifier (encryption key) Coordinate of feature subblocks (encryption key)

14. Patterns and signals to conceal pattern hidden signal pattern hidden signal pattern hidden signal 10000011010102110100 20000112010112210101 30001013011002310110 40001114011012410111 50010015011102511000 60010116011112611001 70011017100002711010 80011118100012811011 90100019100102911100 100100120100113011101 3111110 3211111

15. Backpropagation learning

16. Network 1

17. Network 2

18. Network 3

19. Network 4

20. Network 5

21. Further experiments Can proposed method extract from high pass filtered image or jpeg image ？

22. Network 1

23. Network 2

24. Network 3

25. Network 4

26. Network 5

27. Network 1

28. Network 2

29. Network 3

30. Network 4

31. Network 5

32. Future work Because it relies on the position of feature sub block, it is weak to geometric attacks  Rotation, expansion, shrinking Key sharing has to rely on another security technology

33. Conclusion Information hiding technique without embedding any data into target content by using neural network Ability to collaborate with other information hiding algorithm

34. Thank you

35. Appendix

36. Tradeoffs for information hiding Watermarking (Digital Right Management) Steganography (Covert Channel) Fingerprinting (Integrity check) Capacity (Amount of data to be embedded) Not important Small amount is enough Important More the better Not Important More the better Robustness (tolerance against attack to the container) Important Must not be destroyed Not important Content and hidden data are not related Important Should be weak against alteration Invisibility (transparency of hidden data) Important Should not disturb the content Important Existence should be kept secret Not Important Existence can be informed

37. Three layered perceptron model Three layer model Feed forward model Input function  Sigmoid function Backpropagation learning Input layer Hidden layer Output layer

38. Sigmoid function Input function for multi-layered perceptron model sigmoid = look like letter of S x y

39. Selection of feature values ８ ８ Feature subblock Has DC value and various values of AC (low, middle, high)

40. number of hidden neuron=10 threshold=0.05

41. number of hidden neuron=10 threshold=0.1

42. number of hidden neuron=20 threshold=0.1