Steganalysis with Streamwise Feature Selection

1. Steganalysis with Streamwise Feature Selection Steven D. Baker University of Virginia sdb7e@cs.virginia.edu

2. Steganography: An Example “Hello, I am the amazing Mr. Moulin!” “Hello, I am the amazing Mr. Moulin…” Original Image

3. Motivation • Catch bad people trying to communicate in secret • Catch good people trying to communicate in secret? • Research opportunities: • Improve detection • Disrupt secret communication without harming legitimate image sharing • Improve theoretical guarantees

4. Robustness Secrecy Rate Triangle of Peril Target region Detectable Useless

5. Theoretical work in Steganography • Complexity theory • Provably Secure Steganography [Hopper et al.] • Information theory • An Information-Theoretic Model for Steganography [Cachin] • Perfectly Secure Steganography [Wang and Moulin] • Basic conclusion: perfect security means useless rate • No available, practical algorithm allows smooth adjustment of rate, robustness, and secrecy*

6. PDF Characteristic Function Wavelet transform Training Classification Moments (Feature generation) Feature Selection Method of Wang and Moulin Steg and cover images

7. Let Intel do the work • Can we combine existing features to form useful new features? • Have a computer separate the useless features from the good ones • Do this in a suboptimal but very fast way, so that you can evaluate loads of features, more than there are observations • Streamwise Feature Selection [Zhou et al.]

8. ? Feature generation/selection(Alpha-investing) Add feature Generate feature Reject feature Decrease wealth Increase wealth More features, more time?

9. Feature generation • Pair-wise ratios • Pair-wise differences • Principal Component Analysis • Untested possibilities: • Log • Square root • Nonsmooth Nonnegative Matrix Factorization

10. Experimental data (Shi et al.) • CorelDraw images • ~1000 images • 78 Original features • > 6000 generated features • Steganographic techniques • Cox et al. (SS) • Piva et al. (SS) • Huang et al. (SS) • Generic QIM • Generic LSB

11. Results: ROC

12. Results: Model Complexity

13. Results: AUC

14. Questions?