1 / 42

Spectrum recycling: salvaging analog spectral waste

Spectrum recycling: salvaging analog spectral waste. Kannan Ramchandran EECS Dept. University of California at Berkeley. kannanr@eecs.berkeley.edu http://www.eecs.berkeley.edu/~kannanr. Motivation:. Legacy analog systems can be spectrally v. wasteful AM/FM radio

Télécharger la présentation

Spectrum recycling: salvaging analog spectral waste

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Spectrum recycling: salvaging analog spectral waste Kannan RamchandranEECS Dept. University of California at Berkeley kannanr@eecs.berkeley.edu http://www.eecs.berkeley.edu/~kannanr

  2. Motivation: • Legacy analog systems can be spectrally v. wasteful • AM/FM radio • Broadcast and cable (analog) TV • NTSC TV is 6 MHz. per channel • Digital NTSC-quality ~ 1-2 Mb/s (MPEG) • Analog system spectral efficiency < 0.33 bits/sec/Hz. • Digital spectral efficiency of DTV ~ 4 b/s/Hz. (use 16 QAM) • What’s the problem then? let’s nuke analog… let’s get digital! University of California, Berkeley

  3. Motivation (contd.): • Not so fast… • Radio and television are legacy systems: millions of analog TV’s and AM/FM radios… • Digital simulcast is current FCC solution • Duplicate identical content with extra digital spectrum • Switchover to all-digital mandated by 2006 • Not likely to happen. • Analog systems are here to stay, at least for a while. • Question:Are we stuck then with this spectral inefficiency till we turn the digital corner? • Answer:Not necessarily. University of California, Berkeley

  4. Motivation (contd.): • Can “steal” analog spectrum to do digital transmission • Fully backward-compatible with legacy analog system: • No need to touch existing analog receivers • Digital upgrade system will have superior quality • Can control the switchover from analog to digital • Idea is to “embed” the digital data in the analog signal • Similar in concept to data-hiding/watermarking • Data embedding framework known in theory as channel coding with side-information (CCSI) University of California, Berkeley

  5. ^ X Motivation: Spectrum reuse AM/FM/TV broadcast Legacy receiver X Transmitter Data Embedder Digital Upgrader Data Digital Music/TV Extra data University of California, Berkeley

  6. Motivation (contd.): • Question:How much do we lose in terms of the digital quality due to backward compatibility with analog system? • Answer:Nothing, in some cases, (in theory of course!)… • host is i.i.d. Gaussian signal and channel is AWGN • Analog transmission is actually optimal (analog perf.= digital perf.) • Digital embedding “corrupts” analog system – takes away quality • Digital upgrade system at receiver – fully restores lost quality due to embedding • All-digital system instrumented from scratch cannot do better! • Question:Then why bother with digital systems if analog transmission is optimal? University of California, Berkeley

  7. Motivation (contd.): • Answer: Get real: real-world signals are not i.i.d. Gaussian! • Considerable amount of memory (correlation) • Audio, image, video, speech, text…. • Analog systems ignore the correlation: • no easy way to do analog compression! • Digital systems are much more efficient: • Can pack ~10 NTSC digital channels in the place of 1 analog NTSC channel (and cable companies do!) • So, digital data embedding can allow for: • Simultaneous analog/digital broadcast • No need for digital simulcast on separate spectrum…can use the same analog spectrum! • Analog spectral waste can be recycled seamlessly! University of California, Berkeley

  8. Roadmap for rest of talk • Overview of data embedding: channel coding with side information (CCSI) • Dual of distributed source coding (DISCUS) • Practical examples of data-hiding systems • Data-embedding in images • Data-embedding in audio: toy demo to show power • Other applications and future directions University of California, Berkeley

  9. Data Hiding (Watermarking) • Embedding information in a signal: covert data/ authentication signature • Needs to be minimally perturb host signal (power constraint on the “watermark” added on the signal) • Existing system should be minimally disturbed • Need to be robust to natural and man-made sources of interference • The intended receiver should be able to recover the data/ watermark without the aid of the host signal University of California, Berkeley

  10. Data hiding: channel coding with side info. at the receiver Channel (watermark) (watermark msg.) Decoder Encoder N (attacker) S (host signal) Data Hiding/Embedding Problem • The encoder has access to information S related to the statistical nature of the channel • X is the transmitted signal over the channel University of California, Berkeley

  11. Example: Channel ^ M Y X M + + Decoder Encoder S N Capacity: 1/2 log (1 +X/(S+N)) Example: Channel ^ M Y X M + + Decoder Encoder S N “Writing on dirty paper”: Costa (1982) Capacity: 1/2 log (1 +X/N) independent of strength of S! University of California, Berkeley

  12. CCSI: illustrative example Binary data-embedding/watermarking • Consider a 3-bit host signal S (e.g. binary fax) • Desired to embed data in the host • Max. allowed distortion between S and embedded host X: • Clean channel (no attack) model: received signal Y=X. Case: 1:Both encoder and decoder have access to host signal S: 00  01  10  11  000 001 010 100 4 messages can be embedded: select one of 4 “legal” embedding patterns Encoder outputs X=S+e (mod 2) Decoder receives Y=X and recovers e by: e=S+X (mod 2) University of California, Berkeley

  13. Coset-3 0 0 0 1 1 1 0 0 1 1 1 0 0 1 0 1 0 1 1 0 0 0 1 1 Coset-1 Coset-4 Coset-2 Case 2: When encoder alone knows the host S. Q: Can we still embed 2 bits of information in the S while satisfying distortion constraint between S and X? A: Yes. • Codebook: partition U into 4 cosets • Each of 4 messages indexes a coset in U. • Encoder “perturbs” S to • closest entry X in desired coset of U: • Decoder receives Y=X and • declares coset index of Y as message sent. Messages index one of 4 cosets of U: (10) (00) (11) Example: S=011, m=01; X=001 (off in <= 1 bit) 01 University of California, Berkeley

  14. General encoder and decoder structure for CCSI: DECODER ENCODER Decode Y in the composite channel code and declare the coset containing it as the message Find the coset ‘g’ with the given index Find a codeword, U in coset ‘g’, compatible with S and send X, a function of U and S. X g Y M ^ M Channel S University of California, Berkeley

  15. X X X Codeword Sphere X ENCODING/DECODING - Coset 1 X - Coset 2 - Coset 3 - Side Info - Received Signal Received Signal Sphere (within scale factor) Side-Info Sphere (within scale factor) Assume signal and channel are Gaussian, iid University of California, Berkeley

  16. There is a fundamental “duality” between • CCSI and the problem of distributed coding • (SCSI: source coding with side information) • Encoder and decoder can be interchanged functionally • Allows cross-leveraging of progress between the data • embedding problem and the problem of distributed coding • (DISCUS) used in sensor networks! University of California, Berkeley

  17. ^ Y X Encoder Decoder X Y X Distributed Source coding:(source coding with side information): • The encoder needs to compress the source X. • The decoder has access to correlated side • information Y. • Encoder knows only H(X|Y). Information theory:X can be compressed (in some cases) at a rate equal to that when the encoder too has access to Y (Slepian-Wolf ’72) University of California, Berkeley

  18. DISCUS: source coding with side info. at the Rx X Y Encoder Decoder Channel ^ M X M Encoder Decoder + + S N Duality with channel coding with side info. • Encoder knows some information regarding channel S (not available at decoder) • X transmitted over channel: studied by Gel’fand/Pinsker, Heegard/El Gamal, Costa • Can be applied to “blind” watermarking/data-hiding: host signal available at encoder only. Capacity independent of strength of host signal! University of California, Berkeley

  19. Y Source coding with side information: Illustrative Example ( binary case): Let X and Y be length-3 binary data (equally likely), with the correlation: Hamming distance between X and Y is at most 1. Example: When X=[0 1 0], Y can equally likely be [0 1 0], [0 1 1], [0 0 0], [1 1 0]. ^ • X and Y are correlated. • Y is available at • encoder and decoder. X X Decoder Encoder SYSTEM-1 0 0 0 0 0 1 0 1 0 1 0 0 Need 2 bits to index this. X+Y= University of California, Berkeley

  20. 1 0 0 0 1 1 0 0 0 1 1 1 0 1 0 1 0 1 0 0 1 1 1 0 Coset-1 Coset-3 Coset-4 Coset-2 ^ X X • X and Y are correlated. • Y is available to only • the decoder. Decoder Encoder Y SYSTEM-2 What is the best one can do? The answer is still 2 bits. How? • The Encoder sends the index • of the coset containing X. • The Decoder with this • information and the • knowledge of Y, reconstructs • X without error. University of California, Berkeley

  21. ^ ^ X X X X noisy host source 0 0 0 1 1 1 0 1 0 1 0 1 (00) (01) (10) (11) 1 0 0 0 1 1 0 0 1 1 1 0 Duality:SCSI/CCSI encoder/decoder can be swapped! (010) (10) (10) (010) M: coset index DISCUS Encoder M DISCUS Decoder reconst. S (correlated source) Distributed compression (SCSI) (011) (010) (010) (10) Data-hiding Encoder (10) M: data to be embedded Data-hiding Decoder M embedded host recovered data S (host) (011) Data embedding (CCSI) University of California, Berkeley

  22. X X 1 0 0 0 1 1 0 0 1 1 1 0 0 1 0 1 0 1 0 0 0 1 1 1 X Codeword Sphere X Data-embedding Code Constructions • Want codebook with property that it can be partitioned into “sub-codebooks” (Chou, Pradhan, Ramchandran ’00) • In general, lattices and trellises good (geometrically uniform) • Digital data can then be drawn from a set of “labels” that have a one-to-one correspondence with the “sub-codebooks”. Coset-1 Coset-3 (00) (10) Coset-2 Coset-4 (11) (01) University of California, Berkeley

  23. Data Data Hiding Encoder Rate n/m Rate k/n Host Code Constructions • Consider G0 / 2Zn/ G1embedded coset codes • Framework allows us to partition state-of-the-art channel codes (e.g., turbo codes, TCM) into state-of-the-art source codes (e.g., TCQ) • Our formulation performs near capacity! University of California, Berkeley

  24. Code Constructions (Trellis) • TCM/TCQ encoder Data, d, determines the rate-k/m code to use E[d2] <= X Viterbi Algorithm Rate – k/m code Side Information, S a To Channel + 1-a University of California, Berkeley

  25. Code Constructions (Trellis) • TCM/TCQ decoder Viterbi Algorithm Rate – n/m code Codebook g d’ From Channel (X+S+Z) Calculate Syndrome University of California, Berkeley

  26. Code Constructions (Turbo) • Can extend trellis framework to include turbo codes (channel code is similar to TTCM of Robertson et. al)! Data Rate n/m Rate k/n Data Hiding Encoder Rate n/m p -1 p Side Information University of California, Berkeley

  27. Code Constructions (Turbo) • TTCM/TCQ encoder Side Information, S 1-a a Viterbi Algorithm Rate – k/m code Data, d Rate n/m E[d2] <= X Constellation Mapper + Rate n/m p -1 p To Channel University of California, Berkeley

  28. Code Constructions (Turbo) • TTCM/TCQ decoder From Channel, Y=X+S+Z P(y|gu) MAP + 1 - d’ Calculate Syndrome p p P(y|gu) MAP 1 - Hard Decision + p -1 p -1 University of California, Berkeley

  29. Simulation Results • We use a rate-2/3 convolutional code in concatenation with a rate-3/4 convolutional code for both the TCM/TCQ construction and the TTCM/TCQ construction. (Convolutional codes are constraint-length 4 Ungerboeck codes.) • Assume side-information (S) is i.i.d. Gaussian and Z is also i.i.d. Gaussian: S can be arbitrarily large and can be arbitrary • Embedding rate is 1 bit/sample University of California, Berkeley

  30. Results • At 1 bit/sample, Capacity = 4.77 dB:(C=1/2 log (1 + P/N) regardless of interference strength of side-information S. • Shannon limit if you ignore that S is available at encoder: • C= ½ log(1 + P/(S+N))  If S/N ~ 12 dB, Eb/No  17 dB (gap is 12.23 dB) More recent results (< 2 dB) (< 3.5 dB) 2.72 dB 4.5-5.5 dB University of California, Berkeley

  31. Image Watermarking • Case: Signal (S) is the “Lena” image and the attack is JPEG compression. • Embedding Rate: 1/64 Bits/Sample • Probability of decoding error < 10-5 University of California, Berkeley

  32. Image watermarking: simulation results • Example of robustness of watermark to lossy compression Watermarked image(SDR = 42.22 dB) Original image Can withstand attack up to 32.07 dB (JPEG Q=25%) and yet perfectly embed (with BER < 10-7) up to 4 Kbits of watermarking data in a 512x512 image. University of California, Berkeley

  33. Data Audio Encoded Audio Wavelet Decomposition Coset Code Perceptual Model STFT Audio Data Hiding • Data-hiding capacity can be perceptually optimized • Attractive for legacy systems like FM radio/NTSC TV • Practically possible to hide over 150 kbps in CD quality • audio (noiseless channel) or ~ 45 kbps (14.5 dB SNR channel) with no perceptual degradation • (Chou & Ramchandran ICASSP ’01) University of California, Berkeley

  34. Audio Data Hiding • Model audio coefficients with vector Gaussian (or generalized Gaussian) distribution. • Capacity: C = S Ci • Ci = ½ log(1+Di1/Di2) where Di1 is distortion variance and • Di2 is the channel noise variance University of California, Berkeley

  35. Audio Data Hiding • The amount of quantization noise allowed is determined by the perceptual mask. • Data specifies path in tree • Nodes correspond to source code • Side info needed for depth of tree • In general, we can use a code C0/C1/ … /Cn • Coset codes (Forney) provide nice constructions. Z 0 1 2Z+1 2Z 1 0 0 1 4Z 4Z+2 4Z+1 4Z+3 . . . * * * * * * * * * * * * * * University of California, Berkeley

  36. Audio Data Hiding • Can use a composite trellis code and divide it into multi-stage trellises (Chou et. al. ICIP’ 00) to provide a good channel code and good source codes! • With a good channel code, one can hide data while being robust to channel noise University of California, Berkeley

  37. Applications • Audio data hiding over analog communication channels Data Analog Audio DATA HIDING D/A A/D Data Analog Receiver Analog Audio Channel Digital Receiver A/D University of California, Berkeley

  38. Design and Simulation results: • Audio data sampled at 44.1 kHz (CD Quality) • 6 Stage scalar quantizer (with and without FEC) • Without FEC can hide around 150 kbps (of course this is with disregard to the channel) • With BCH codes can hide 42.7 kbps (and transmit reliably over channels with 14.5 dB SNR). University of California, Berkeley

  39. Design and Simulation results (cont.): • With better codes (i.e., Trellis codes, etc) should be able to perform even better! • Original Audio File (44.1kHz,14.99 sec.) • Audio File (44.1kHz, 14.99 sec.)with 154.7 kbps (2.32 Mbits total) of data hidden in! University of California, Berkeley

  40. Channel p(y|x,s) Enc 1 Dec 1 Enc N Dec N New user New user Big picture: new constructive way to do multiuser communication • Can add more user(s) by “piggybacking” signal on compound signal of other users: minimal obtrusion on other users: fully backward compatible with existing receivers! • Constructive way to do broadcast (optimal theoretical way!) University of California, Berkeley

  41. Other communication system applications • Multi-antenna broadcast (BS to mobiles) • Embedding users’ information inside one another’s signals is information-theoretically optimal. • Downlink capacity can be increased. • ISI cancellation (improved precoding) • Treat ISI noise as “side information” • DSL cross-channel interference • CO hub to residential units is a broadcast channel • Can treat cross-channel interference as side-information that is deterministically known (can be completely removed in theory!) University of California, Berkeley

  42. Data-hiding idea is very powerful and can be • applied to the original problem of spectrum • recycling of wasteful analog bandwidth • Challenges are many-fold: theoretical, algorithmic, • implementational and system-level. • Target specific applications of interest • BWRC is perfect place to make a lot of this happen! Conclusions and future directions University of California, Berkeley

More Related