Cryptagram. Photo Privacy for Online Social Media Matt Tierney, Ian Spiro Christoph Bregler, Lakshmi Subramanian

Transcription

1 Cryptagram Photo Privacy for Online Social Media Matt Tierney, Ian Spiro Christoph Bregler, Lakshmi Subramanian Courant Institute, NYU

2 Photo Privacy Problem

3 Photo Privacy Problem

4 Photo Privacy Problem

5 Photo Privacy Problem OSN?

6 Cryptagram empowers users with usable end-to-end photo privacy in OSNs

7 Photo Privacy Problem Given an image I and secret key k; image embedding protocol G(), G (); lossy compression transformation T(), T (); and quality q for T():

8 Photo Privacy Problem

9 Photo Privacy Problem Client-Side OSN

10 Photo Privacy Problem Client-Side E(, k) OSN

11 Photo Privacy Problem E(, k) Client-Side m = OSN

12 Photo Privacy Problem E(, k) Client-Side G m = OSN

13 Photo Privacy Problem E(, k) Client-Side G m = OSN T

14 Photo Privacy Problem E(, k) Client-Side G m = OSN T T

15 Photo Privacy Problem E(, k) Client-Side G m = m = G OSN T T

16 Photo Privacy Problem E(, k) Client-Side G m = E (k, m = G OSN T T

17 q,p-recoverability Given a message m (a sequence of bits); embedding protocol G(), G (); lossy compression transformation T(), T (); and quality q for T(): With what probability is m encoded and decoded to the same value? Given a quality value set by as OSN, what probabilistic guarantees of data integrity does an embedding protocol give?

18 q,p-recoverability Given a message m (a sequence of bits); embedding protocol G(), G (); lossy compression transformation T(), T (); and quality q for T(): With what probability is m encoded and decoded to the same value? Given a quality value set by as OSN, what probabilistic guarantees of data integrity does an embedding protocol give?

19 q,p-recoverability Given a message m (a sequence of bits); embedding protocol G(), G (); lossy compression transformation T(), T (); and quality q for T(): With what probability is m encoded and decoded to the same value? Given a quality value set by as OSN, what probabilistic guarantees of data integrity does an embedding protocol give?

20 Design Requirements Data confidentiality for end-users OSN-shared photos Probabilistic data integrity for lossy images Usable solution that seamlessly integrates with the OSN Widely deployable and applicable protocols and solution Efficient packing of embedded data to retain as much of the original image

21 Cryptagram Encoding

22 Cryptagram Encoding

23 Cryptagram Encoding S e c u r i t y

24 base64 index Cryptagram Encoding S e c u r i t y

25 base64 index Cryptagram Encoding S e c u r i t y

26 base64 index Cryptagram Encoding S e c u r i t y

27 base64 index Cryptagram Encoding S e c u B = r i t y

28 base64 index Cryptagram Encoding S e c u B = r i t y

29 base64 index Cryptagram Encoding S e c u B = r i t y

30 base64 index Cryptagram Encoding S e c u B = r i t y =

31 base64 index Cryptagram Encoding S e c u B = r i t y =

32 Cryptagram Encoding B = =

33 Cryptagram Encoding B = =

34 Cryptagram Encoding Cryptagram Pixel Block (CPB) B = =

35 Cryptagram Encoding Cryptagram Pixel Block (CPB) B = =

36 Cryptagram Decoding

37 Cryptagram Decoding

38 Cryptagram Decoding

39 Cryptagram Decoding

40 Cryptagram Decoding S e c u r i t y

41 Luminance-Only q,p-recoverability enlightens us to how destructive lossy compression is to our codec; that is, how often pixels are too perturbed to decode correctly

42 Luminance-Only Facebook q=74 We can choose protocols based on desired packing efficiency, filesize expansion, and the quality threshold imposed by the OSN

43 Luminance and ECC Facebook q=74 Using RS(255, 223) we can recover from at most 6% errors and therefore reduce the necessary threshold along q,p-recoverability curves to determine utility of an embedding protocol

44 Summary of Protocol Efficiency Without ECC With ECC (RS(255,223)) lum B q,100-rec (quality) Efficiency (bpp) q,94-rec (quality) Effective Efficiency (bpp) Hex Oct Quad Bin < 20 1 < We have q,94-recoverability for Facebook with an effective efficiency of 2.62 bits per pixel using an embedding protocol with ECC

45 Summary of Protocol Efficiency Quality File size Expansion A user can choose to balance Cryptagram quality, target q,p-recoverability, and file size

46 Implementation Open-sourced code for various platforms Chrome and Firefox Extensions (JavaScript) Drag-and-drop Web-based Offline Encoder (JS) ios (Obj-C) and Android (Java) Testing frameworks (C++ and Python) 400+ Chrome extensions active users 373 IRB-approved users sending high-level data 3,300 decrypted Cryptagrams

47 Future Directions

48 Future Directions Compressing the compressed

49 Cryptagram Summary Problem: Transmit images through social networks while minimizing risks to privacy Solution: Cryptagram uses spatial domain embedding codecs with sufficient q,p- Recoverability Usable platform (client-side encoder; Facebook, G+, etc. decoder) for ongoing user study

50 Cryptagram empowers users with usable end-to-end photo privacy in OSNs

51 Cryptagram Demo

52 Cryptagram Demo

53 Cryptagram Photo Privacy for Online Social Media Matt Tierney, Ian Spiro Christoph Bregler, Lakshmi Subramanian Courant Institute, NYU

54

55 Four Horseman of the Infocalypse [Cypherpunk FAQ] Cypherpunk FAQ "How will privacy and anonymity be attacked?" like so many other "computer hacker" items, as a tool for the "Four Horsemen": drug-dealers, money-launderers, terrorists, and pedophiles.

56 Chrominance and Luminance Chrominance is too inefficient to be useful for efficient packing goal But what about usability?

57 Usable Identifiers We can embed icons in the chrominance spaces (without impacting luminance encoding) to enable users to identify images to one another

58 Handling Partial Failures Dividing the input image into distinct chunks, we encrypt and embed each chunk separately Chunk order is randomized, with headers containing sequence numbers

59 Related Work Project Technique Unresolved Challenges P3 [NSDI 13] Split JPEG s DCT coefficients into two photos: secret vs public. Use multiple services to store different photos Susceptible to face identification Need another third-party service account JPEG-centric Closed source - converting to startup Robust Image Obfuscation [SPIE 12] (1) Tile shuffling (2) Masking Polynomial-time descrambling algorithms defeat technique (1) Unclear bijective permutation function design Masking with pseudo-random pattern for grayscale Offline solution Chaotic DCT Coefficient Shuffling [ICME 09] Shuffle like-frequency DCT coefficients using logistic maps (chaos theory) Security with chaos-based crypto is broken Do not examine compression artifacts in depth JPEG-centric Image size cannot be a secret

60 Related Work Project Data Confidentiality Probabilistic Data Integrity Seamlessly Usable Wide Deployability and Applicability Efficient Packing P3 [NSDI 13] Robust Image Obfuscation [SPIE 12] Chaotic DCT Coefficient Shuffling [ICME 09]

61 Related Work Project Data Confidentiality Probabilistic Data Integrity Seamlessly Usable Wide Deployability and Applicability Efficient Packing P3 [NSDI 13] x-pire [arxiv 11] Robust Image Obfuscation [SPIE 12] Chaotic DCT Coefficient Shuffling [ICME 2009]

62 JPEG Review JPEG compression operates on luminance, chrominance blue, and chrominance red (transformation from red, green, blue) JPEG compresses luminance and chrominance differently

63 JPEG Review JPEG compression operates on luminance, chrominance blue, and chrominance red (transformation from red, green, blue) JPEG compresses luminance and chrominance differently

64 8 JPEG Review 8 JPEG compression operates on luminance, chrominance blue, and chrominance red (transformation from red, green, blue) JPEG compresses luminance and chrominance differently

65 8 JPEG Review 8 JPEG compression operates on luminance, chrominance blue, and chrominance red (transformation from red, green, blue) JPEG compresses luminance and chrominance differently

66 8 JPEG Review 8 Lossy JPEG compression operates on luminance, chrominance blue, and chrominance red (transformation from red, green, blue) JPEG compresses luminance and chrominance differently

67 DCT 8 JPEG Review 8 Lossy JPEG compression operates on luminance, chrominance blue, and chrominance red (transformation from red, green, blue) JPEG compresses luminance and chrominance differently

68 DCT 8 JPEG Review 8 Lum Q Chrom Q Lossy Lossy JPEG compression operates on luminance, chrominance blue, and chrominance red (transformation from red, green, blue) JPEG compresses luminance and chrominance differently

69 DCT 8 JPEG Review 8 Lum Q Entropy Encoding Chrom Q and Write to Disk Lossy Lossy JPEG compression operates on luminance, chrominance blue, and chrominance red (transformation from red, green, blue) JPEG compresses luminance and chrominance differently

70 Contributions Efficient top-down, widely applicable embedding protocols For lossy compression, we have a quantified notion of probabilistic data integrity Multi-variate evaluation for different protocol designs and image codecs Live deployment with real users