1. 1 Detecting Hidden Messages using higher-order stats and SVMs Siwei Lyu and Hany Farid

2. 2 Do higher order statistics detect hidden messages? ➢ Why higher order? ➢ What features? ➢ What Scale? ➢ How to compute? ➢ How to “learn” what features matter in classification

3. 3 Wavelet Features

4. 4

5. 5

6. 6

7. 7

8. 8

9. 9 Main Hall Dartmouth

10. 10 Wavelet decomposition

11. 11

12. 12 Predicting across layers

13. 13 Horizontal Prediction

14. 14

15. 15

16. 16 The 72 features...

17. 17 Approach ➢ Using 72 features, and a large training set ➢ Generate features for clean images: Negative examples ➢ Generate features for different steg algorithms on that each image: Positive examples ➢ Build classifier using the positive and negative rtaining examples. ➢ Test on images not used in training. ➢ Test on steg algorithms not used in training

18. 18 Experimentation

19. 19

20. 20

21. 21

22. 22

23. 23 SVM introduction

24. 24 “Hyper plane separation”

25. 25 SVM Continued ➢ Solve through Lagrange multipliers. ➢ Can do non-linear by lifting into embedding space. (Tricky) but because only need inner products, its not too expensive. ➢ Use library, SVM is just a good “learning” tool ➢ (paper should never have included this, should say go read the references).

26. 26

27. 27

28. 28