Dec 26, 2015
Multimedia: crypto is not enough
19/04/23 | pag. 2
Crypto Problem for Multimedia
Once decrypted, the multimedia data becomes vulnerable for distribution.
Cryptography 19/04/23 | pag. 3
Watermarking, which is imperceptible message embedding within the work, complements encryption.
Watermarking
Cryptography 19/04/23 | pag. 4
Channel
Detect Watermark
Embed Watermark
Modern Applications
• Copyright Protection• Fingerprinting or Traitor Tracing• Content Authentication• Media Forensics• Steganography• Database Annotation• Device Control• …
Cryptography 19/04/23 | pag. 5
Basic Technical Issues (1)
• Transparency (fidelity): embedding information should not cause perceptual degradation of host signal
• Payload (bit-rate): number of bits that can be embedded in signal
• Robustness: refers to ability of embedding algorithms to survive common signal processing operations (compression, filtering, noise, cropping, insertions,…)
Cryptography 19/04/23 | pag. 6
Basic Technical Issues (2)
• Security: refers to ability of an adversary to crack information hiding code and design devastating attack to wipe out hidden information (e.g. recovery of host signal which contains no trace of message).
• Detectability: is the fact of information embedding a secret (cfr. Steganography)?
There exist fundamental trade-offs between transparency, payload or bit rate, robustness, security and detectability.
Cryptography 19/04/23 | pag. 7
System issues
• Does decoder have full, partial, or no knowledge of host signal?
Availability of side information generally improves detection BUT introduces communication and storage burden.
• What security level is needed?• What are the attacker’s computational resources?• Does attacker have repeated access to decoder?
Cryptography 19/04/23 | pag. 8
Standards and Benchmarking
• So far: NO foolproof watermarking due to lag between theory and practice.
• Examples: • Music industry used watermarking to protect digital music• Watermarking as part of international MPEG-4 video standard
• Recent research groups have developed benchmark tools:• Stirmark• European Certimark• WET project at Purdue university• …
Cryptography 19/04/23 | pag. 9
Mathematical Models
Encoders and decoders:
• Encoding function
x=f(s,m,k)
• Decoder function
m’=g(y,k)
*source:Data-Hiding codes (Moulin)Cryptography 19/04/23 | pag. 10
Mathematical Models
Attacks:
*source:Data-Hiding codes (Moulin)Cryptography 19/04/23 | pag. 11
Mathematical Models
• Distortion (to characterize perceptual closeness):
Squared Euclidian metric (audio, grayscale images (PSNR),…):
Hamming distance (binary images):
Fail to capture complexities of human perception Perceptual models are needed:
• Popular example is Watson’s metric
• SSIM
Hd ( , ) | : | , if {0,1}n nn s x S X s x
Cryptography 19/04/23 | pag. 12
Early Work
1990’s:• First papers on data hiding appeared• Least significant bit (LSB) techniques• Problem: not robust against noise
1995-1998:• Spread-spectrum modulation (SSM) codes• More robust
1998-…:• Quantization Index Modulation• Very good performance
Cryptography 19/04/23 | pag. 13
LSB codes
• Host signal s = {s1, s2, … ,sN}
• Each si uses b bits (integer value between 0 and 2b-1)
e.g.: si = 65 = (01000001)
• LSB plane is length-N binary sequence made of all LSB’s• Information rate is 1 bit per sample.• Payload can be increased by replacing second LSB
increase of embedding distortion!
Very weak against noise
Cryptography 19/04/23 | pag. 14
LSB codes
19-04-23
Channel
Detection
Embedding
1 bitplane
19-04-23
2 bitplanes
19-04-23
3 bitplanes
19-04-23
4 bitplanes
19-04-23
5 bitplanes
19-04-23
6 bitplanes
19-04-23
7 bitplanes
19-04-23
8 bitplanes
19-04-23
LSB
Cryptography 19/04/23 | pag. 24
LSB
Cryptography 19/04/23 | pag. 25
Spread-Spectrum Codes
• Watermarking problem ≈ communication problem with a jammer:
Apply techniques from communications domainSpread-Spectrum Modulation!
Jamming problem:• Classic radio/TV transmitter sends signal in relatively
narrow frequency band• Inappropriate with jammer that allocates all power to that
particular band of frequencies
Cryptography 19/04/23 | pag. 26
Spread-Spectrum Codes
SSM-system: • Allocates secret sequences (with broad frequency
spectrum) to transmitter.• Transmitter sends data by modulating these sequences.• Receiver demodulates data using filter matched to secret
sequences.
Jammer must spread power over broad frequency range but only small fraction of it will have an effect on communication performance.
Cryptography 19/04/23 | pag. 27
Spread-Spectrum Codes
• Sender:
xn = sn + αpn(m,k), 1 ≤ n ≤ N
• Receiver: y = x + w, w noise
• Knows secret key k• Matching of y to all possible
waveforms p(m,k)
*source:Data-Hiding codes (Moulin)Cryptography 19/04/23 | pag. 28
Spread-Spectrum Codes
Detection:• Correlation:
• Informed detector:
*source:Data-Hiding codes (Moulin)Cryptography 19/04/23 | pag. 29
Binning Schemes
• An important information-theoretic technique, which is widely used.
• Especially for blind data hiding, which is related/complementary to the problem of transmission with side information.
Cryptography 19/04/23 | pag. 30
Binning Schemes
• Quantization on source sequence S of length N
• Quantization codebook C of length-N vectors Uj
• Distortion function D
• VQ problem = find vector Uj within codebook C that minimizes distortion D(S, Uj) between observed S and reconstruction vector Uj
• Next, consider M different codebooks Cm (= bins) consisting of length-N vector Um,j within C now we can choose which codebook we want to use!
With M codebooks we can embed k (=log2M) bits
Cryptography 19/04/23 | pag. 31
Example: Binning Schemes
• S = binary sequence of length N = 3 (8 possible sequences).• We want to embed a information into S, producing a new
sequence X.• We must satisfy distortion constraint that X and S differ in 1
bit.• Transmission of X to a receiver which must decode
embedded info without knowledge of original host data S.
Question 1: how many bits can we embed in S?Question 2: How can we design appropriate scheme?
Cryptography 19/04/23 | pag. 32
Binning Schemes example
• Under distortion constraint, S can be modified in 4 ways:• S=010 => X = {000, 110, 010, 011} 2 bits of info can be embedded
• Consider partition of eight possible X into 4 bins.• e.g.: S = 010 and Hamming distance of one
Cryptography 19/04/23 | pag. 33
Binning Schemes example
Embedding 1-bit message into 7-bit length sequences S.
2 bins allow us to modify 3 bits of S.
In this way, it resists the noise with Hamming weight at most 1.
Cryptography 19/04/23 | pag. 34
QIM
• Quantization-Based Codes have been introduced in 1999, by Chen and Wornell, known as Dither Modulation or Quantization Index Modulation (QIM) codes.
• QIM is a binning scheme.
• Theses methods embed signal-dependent watermarks using quantization techniques.
Cryptography 19/04/23 | pag. 35
Scalar QIM
Embedding one bit in a real-valued sample:
Here we have 1-bit message m {0,1}.∈A scalar, uniform quantizer Q(s) with step size is defined △
as Q(s) = [s/ ].△ △We use the function Q(s) to generate two new dithered
quantizers.
Cryptography 19/04/23 | pag. 36
Scalar QIM
The two new dithered Quantizers Q0 and Q1 are shown as sets of circles and crosses on the real line.
Cryptography 19/04/23 | pag. 37
Original QIM
The noisy signal Y = X (Marked signal)+ W.
• The QIM decoder is a minimum-distance decoder. It finds the quantizer point closest to Y.
• The quantization errors are uniformly distributed over
[- /2, /2].△ △
• This scheme works perfectly, if |W| < /4.△
Cryptography 19/04/23 | pag. 38
Distortion-Compensated Scalar QIM
• A binning scheme with some error protection against noise (exceeds /4).△
• The embedding function is defined as:
Cryptography 19/04/23 | pag. 39
Distortion-Compensated Scalar QIM
Cryptography 19/04/23 | pag. 40
Other QIM Schemes
• Sparse QIM
• Lattice QIM – Another extension of Scalar QIM to the vector case.– Replace the scalar quantizer with a L-dimensional
VQ quantizer
Cryptography 19/04/23 | pag. 41
Data Hiding in Images
Quantization-based codes are widely used to embed data in images.
Cryptography 19/04/23 | pag. 42
Data Hiding in Images
Lena after data embedding and after attack
Cryptography 19/04/23 | pag. 43
Data Hiding in Images
• 6301 bits was embedded in Lena and tampered with the marked image in various ways• Cropping• Resizing• Substitutions• Compression• Moderate noise levels
• All 6301 bits could be successfully decoded.
Cryptography 19/04/23 | pag. 44
Conclusion
• The information-theoretic concept of binning known as the best methods, when the host signal is unavailable to the receiver (blind data hiding).
• Practical binning schemes show good performance under noise attacks.
• Spread-spectrum techniques are popular, but have limitations for blind data hiding.
Cryptography 19/04/23 | pag. 45
Thank you!
… the vast majority of security failures occur at the level of implementation …
Cryptography 19/04/23 | pag. 46