Priming for Steganography

Presentation about an experiment done by myself and Derek Kern to determine if filters can be used to preprocess images prior to steganography embedding to reduce efficacy of steganalysis.

Transcript

1. Priming for Steganography Undermining Steganalysis through Artificial Noise Introduction Derek Kern Rod Hilton
2. In World War II, a group of allied troops had been captured by German forces. A rescue team was sent in, only to discover the compound where they were being held was outfitted with state-of- the-art motion detectors. There was no way to infiltrate without setting off the alarms. The allies came up with a clever plan. One night they tossed rocks over the fence until the alarms were set off. The Germans searched for hours, only to decide it must have been a false alarm. The next night the Allies did the same thing, intentionally setting off the alarm. Again, the Germans found nothing. After a week of this, the Allies tossed the rocks over the fence and nothing happened. The Germans had concluded their system was faulty and disabled it. The allies raided the compound that night and rescued the prisoners.
3. Sometimes, the best attack on a security system... ...is to attack the confidence people have in it.
4. What can we do to undermine confidence in steganalysis? Maybe we modify images prior to embedding to mislead the detector.
5. How does steganalysis work? We’ve talked a lot about how steganalysis works mechanically, but what’s it really doing, abstractly? Fundamentally, it’s looking for “suspicious noise” in an image
6. Can we add artificial, innocuous “noise” to an image before embedding? Artificial Noise Stego Noise Can the detector tell the difference between the two types of noise?
7. Hypothesis: we can add artificial noise to images prior to embedding to reduce efficacy of steganalysis Classified Cover Classified Stego Actual Cover True Negative False Positive Actual Stego False Negative True Positive
8. Couldn’t you just fill the image with random noise? Yes, but that would be cheating Goal of steganography is to embed in images that look normal
9. We can’t just insert a ton of noise, it has to be a picture a person would send. Filters preserve images while potentially adding “noise” Contrast
10. Start with 1,000 basic images Selected for no Steganography Selected for Steganography Embed message (20% payload) steghide! Steganalyze & generate confusion matrix Apply chosen filter Generate ROC curve for filter stegdetect! Calculate AUC, compare
12. Filter: Random Completely ignores input data, generates random pixels. Useful for comparison purposes
13. Performance: Random 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 AUC: 0.50155475
14. Filter: Baseline Applies no filter, this represents how effective stegdetect is without us messing with the images
15. Performance: Baseline 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Baseline Random Why isn’t this better? AUC: 0.575883735
16. Filter: Noise Adds monochromatic noise to image (look at the rim of the plate), directly trying to insert noise without compromising image
17. Performance: Noise 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Baseline Noise Random Baseline AUC: 0.575883735, Noise AUC: 0.596505775 Pretty close to baseline, in fact slightly higher (stegdetect does better)
18. Filter: Saturate Deepens colors, replacing grayish colors with fuller versions of those colors. Generally increases contrast.
19. Performance: Saturate Baseline AUC: 0.575883735, Saturate AUC: 0.602616 Again, close to baseline (the highest yet) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Baseline Saturate Random
20. Filter: Contrast Increases contrast in the image, shadows become darker, highlights become lighter.
21. Performance: Contrast Baseline AUC: 0.575883735, Contrast AUC: 0.52535131 Contrast weakens stegdetect! 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Baseline Contrast Random
22. Filter: Blur Blurs the image at 6x
23. Performance: Blur Baseline AUC: 0.575883735, Blur AUC: 0.50324375 Blur highly diminishes analyzer power. Close to Random. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Baseline Blur Random
24. Filter: Sigmoidal Increases the contrast of the image using a sigmoidal transfer function without saturating highlights or shadows
25. Performance: Sigmoidal Baseline AUC: 0.575883735, Sigmoidal AUC: 0.53883085 Sigmoidal diminishes the abilities of detection 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Baseline Sigmoidal Random
26. Filter: Emboss Emboss the image with a radius of 1 pixel
27. Performance: Emboss Baseline AUC: 0.575883735, Emboss AUC: 0.46764815 Emboss is more effective at weakening stegdetect than Random! 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Baseline Emboss Random
28. Whoa. stegdetect is over 35,000 lines of C code… Its lead developer is a Google engineer with a PhD in CS… It has been worked on for over a decade… …and embossing an image first makes it perform worse than flipping a coin (if the image is embossed, you’d be better off betting against stegdetect)
30. Filter: Tilt Shift Applies a “tilt shift”-like effect, effectively blurring the top and bottom of the image
31. Performance: Tilt Shift Baseline AUC: 0.575883735, Tilt Shift AUC: 0.659988555 Tilt Shift makes stegdetect much more effective! (Even though Blur makes it less) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Baseline Tiltshift Random
32. Filter: Kelvin Strong peach-colored overlay to the image, desaturating the image substantially so that peach is the dominant color. It also adds a washed-out frame to the edges, making the image look like a deteriorated photograph
33. Performance: Kelvin Baseline AUC: 0.575883735, Kelvin AUC: 0.69150218 Kelvin improves performance of stegdetect more than any other 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Baseline Kelvin Random
34. Filter: Toaster Meant to emulate old Polaroids, this filter gives the image a pinkish glow with increased color saturation. It also adds a vignette, a desaturation around the edges of the images giving the borders a somewhat "burnt" look.
35. Performance: Toaster Baseline AUC: 0.575883735, Toaster AUC: 0.497654305 Toaster greatly diminishes usefulness of stegdetect! 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Baseline Toaster Random
36. Filter: Lomo Emulates "toy cameras" meaning cameras with a plastic lens, which tend to add a soft focus and shadowy border. This is done by increasing the red channel constract substantially and applying a vignette that darkens/burns the edges of the image.
37. Performance: Lomo Baseline AUC: 0.575883735, Lomo AUC: 0.571520055 Lomo is about equal to baseline! 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Baseline Lomo Random
38. Filter: Nashville Emulates old fashion photos from the 80's. Resulting images have a slight peach tint and a frame around it that makes it look like a slide.
39. Performance: Nashville Baseline AUC: 0.575883735, Nashville AUC: 0.500399225 Nashville makes detection more difficult, close to Random 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Baseline Nashville Random
40. Filter: Gotham Converts an image to black and white with high contrast and a blue tint. Meant to emulate pictures taken with a Holga camera and Ilford X2 film.
41. Performance: Gotham Baseline AUC: 0.575883735, Gotham AUC: 0.48454428 Gotham does the best job of weakening the detector of all the Instagram filters 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Baseline Gotham Random
42. Results & Conclusions Filter AUC Kelvin 0.69150218 Tilt Shift 0.659988555 Saturate 0.602616 Noise 0.596505775 Baseline 0.575883735 Lomo 0.571520055 Sigmoidal 0.53883085 Contrast 0.52535131 Blur 0.50324375 Random 0.50155475 Nashville 0.500399225 Toaster 0.497654305 Gotham 0.48454428 Emboss 0.46764815 Improve steganography detection Diminish steganography detection Nullify steganography detection
43. Results & Conclusions Filter AUC Kelvin 0.69150218 Tilt Shift 0.659988555 Saturate 0.602616 Noise 0.596505775 Baseline 0.575883735 Lomo 0.571520055 Sigmoidal 0.53883085 Contrast 0.52535131 Blur 0.50324375 Random 0.50155475 Nashville 0.500399225 Toaster 0.497654305 Gotham 0.48454428 Emboss 0.46764815 •  We can use image filters to introduce noise that undermines steganalysis •  In fact, some filters are even better at doing this than we expected •  Some filters can even be used to assist steganalysis •  Hypothesis confirmed!
44. Artifacts Improved steganography through: •  Use of certain pre-embedding filters •  Avoid other pre-embedding filters Improved steganalysis through: •  Constructing databases with artificial noise to train classifiers/steganalyzers
45. Questions ?