Lecture 6. SG based on noisy channels [4]. CO Detection of SG The peculiarities of the model: - it is alloed to use by attacker only noisy version of SG C (n), - CO can be knon exactly for attacker. Practical applications: - transmission of SG over satellite, mobile or optical fiber channels, - interception SG by attacker over side channels, - simulation of noisy channels. - slightly noised musical files placed on some sites in the Internet. 1
The main idea to design SG based on noisy channels: Camouflage the embedded message by noise of the channel. Attacker s problem in this setting : To distinguish if there exists channel noise only either both channel noise and SG. Simplification in design of SG based on noisy channel in comparison ith conventional case: Description of distribution for channel noise is simpler than description of CO distribution. We consider to types of the channels: 1. Binary symmetric channel ithout memory (BSC).. Continuous channel ith Gaussian hite additive noise. Condition. We ill assume that legal and illegal channel have the same distributions of noise.
1. SG based on BSC. Embedding: C( n) = C( n) Θ b( n) π ( n), n= 1,.. Cn ( ) {,1}; π( n) {,1}, iid..., P( π( n) = 1) = P, P( π( n) = ) = 1 PΘ ( n) = b, n= 1,.. " " - modulo to addition., b (1) After a passing of C (n) over BSC e get: Cʹ ( n) = C( n) ε ( n), n= 1,.. ε( n) {,1}; iid..., P( ε( n) = 1) = P, P( ε( n) = ) = 1 P, P < 1/. Attack in order to detect SG: 1.!C (n) = C! (n) C(n), (C(n) is knon exactly by attacker ) (3). To hypothesis testing: - Bernoulli sequence ith the parameter P, - Bernoulli sequence ith the parameter H :! C (n),n =1,.. H 1 :! C (n),n =1,.. P = P(1 P ) + P (1 P). 1 () 3
Evident testing method : H, if t < τ, H 1, if t τ, (4) here t = H ( C! (n),n =1,.. ), H (X) the Hamming eight of X. some threshold. Performance evaluation of SG detecting is determined by probabilities P fa and P m : τ 1 i i i i Pm = P1(1 P1), Pfa P(1 P), i= 1 i = i= τ i!. i = i!( i)! here τ But the exact calculation by (5) is hard for > 1 3. Thus e ill use the criterion of relative entropy (see Lecture 5). (5) 4
PH ( x) D= D( PH P ) ( )log, H1 = PH x P ( x) x X For BSC model here X = {,1}, e get easily P 1 P D= Plog + (1 P)log. P1 1 P1 Then the bound for P fa and P m (see Lecture 5) is : Pfa 1 Pfa Pfa log + (1 Pfa)log D, 1 P P fa here D is calculated by (6). m H 1 (6) (7) ext e ill find the probability of error P e under extraction of the embedded bit by informed legal decoder. Decoding scheme: Λ = ( C " (n) C(n))π (n) b! ' 1,Λ λ = (, ),Λ < λ n=1 here λ is some threshold, is the number of samples hich ere used to embed one information bit. (8) 5
Substituting (1) and () into (8), e get Λ = Λ b for b = and b = 1, here Then Λ = ε( n) π( n), Λ = ( ε( n) π( n)) π( n). 1 n= 1 n= 1 δ = 1 here (11) H (π) is the Hamming eight of the sequence π(n), n=1,. Substituting (1) and (11) into (1) e get In a similar manner e can obtain that P(/1) = P(/1, s) P( H ( π ) = s), P(/1, s) = P( Λ < / s) = s P (1 P) PH s P P s s j s j 1 λ j= s λ j s s ( ( π ) = ) = (1 ). s s P P P P P s= 1 s j= s λ j s s j s j (/1) = (1 ) (1 ). s s P P P P P s= 1 s j= λ j s s j s j (1/ ) = (1 ) (1 ). (9) (1) (1) (13) (14) 6
Because the sequence π(n) is knon at the decoder it is possible to get P(/1) = P(1/), selecting λ = S/, that gives s s Pe = P = P = P P P P s= s j= s/ j s s j s j (/1) (1/) (1 ) (1 ). We apply Chernoff bound to inner sum in (15) in order to simplify it : s j= s/ s s P j [ ] / (1 P s j ) 4 P (1 P ), j and substituting (16) into (15) e get s Pe P P P P s= s [ ] s (1 ) 4 (1 ). Finally using Binomial eton formula in (17) e have s / (15) (16) (17) Pe P P P+ ( (1 ) 1) 1, K K ( a+ b) = a b. K= K because (18) 7
Optimization problem in design of SG based on noisy channels : Given D (security), P (state of BSC), P e (the desired reliability) it is necessary to find the parameters P and, that provide maximum of the number secure embedded bits m = /. / m Pe ( P(1 P) 1) P 1 +, P 1 P D= Plog + (1 P)log, P1 1 P1 P = P(1 P ) + P (1 P). 1 Example. D=.1, P =.1, P e =.1. Then m=63 for =137391, P = 6.7 1-7. If e take the restriction 1 6, then m=19 for P = 8.63 1-6. (19) 8
. SG based on Gaussian channels. Embedding: b C( n) = C( n) + ( 1) σπ ( n), n= 1,.., π( n) iid... (,1), σ here amplitude coefficient. After a passing of SG C (n) over Gaussian channel e get: Cʹ ( n) = C( n) + ε ( n), n= 1,.., here ε( n) iid... (, σε ). (1) () Attack to detect SG consists in to steps: 1. (С(n) is knon exactly for attacker ).!C (n) = C! (n) C(n),. To hypothesis testing: H :! C (n),n =1,..,i.i.d., (,σ ε ), H 1 :! C (n),n =1,..,i.i.d., (,σ ε +σ ). There exists ell knon testing methods for this model but it is more convenient to apply relative entropy technique for to Gaussian distributions. (3) 9
+ PH ( x) D= D( PH P ) ( )log, H = 1 PH x dx PH ( x) 1 P ( x) (, σ ), P ( x) (, σ + σ ). here After a calculation of the integral in (4) and simple transforms e get : here H ε H ε 1 D= + 1 + η 1,7 ln(1 ) (1 ), η η = σ / σ ε (noise-to-atermark ratio (WR)). For large WR, that provides high security of SGS, e have from (5) D,36. η η Express, from (6) as a function of and D: η.6 / D. (4) (5) (6) (7) 1
ext e find the probability P e under extraction of the embedded bit by legal informed decoder. Decoding scheme. Λ = ( C " (n) C(n))π (n) b! $ 1,Λ = %. &,Λ < n=1 Substituting (1) and () into (8) e get The probability of error P e = P(/1) = P(1/) can be easily found by (9) taking into account that Λ is Gaussian random variable oing Central Limit Theorem: here b Λ= ( ε( n) + ( 1) σ π( n)) π( n). n= 1 E{ Λ} Pe = Q, Var { Λ} t 1 Qx ( ) = e dt. π x (8) (9) (3) 11
After simple calculation of E{Λ} and Var{Λ} e get σ σ P e = Q Q = Q. σ σ η ε σ + ε Substituting (7) into the last equality e obtain ( ) 1/ P = Q 1.9 ( D) / m. e Remark. There is a simple bound for the function Q(x) [35 ]: Qx ( ) exp( x ). Substituting (33) into (3) e get 1/ ( ) P exp.83( D) / m. e Example. Given D=.1, =1 e get P e 3 1-4. (31) (3) (33) (34) Important remark. In Gaussian noisy channel it is possible to embed any number of bits m given any security level D and any reliability P e.(in contrast to BSC case!) 1
General conclusions for SG based on noisy channels. 1. This is unique case here it is possible to provide a secrecy of SG in attack ith knon exactly CO.. A choice of channel model (BSC or Gaussian) depends on the condition in hich place, of telecommunication line is alloed to embed and extract additional information. 3. Design of SG does not require the knoledge of CO statistics. 4. In the case of BSC the number of secure and reliably embedded bits is limited, hereas in the Gaussian case it is unlimited (at the coast of increasing of ) but the embedding rate approaches to zero as. 5. The use of error correcting codes allos to increase slightly the embedding rate but it is kept as before approaching to zero as approaches to infinity on the contrary to conventional communication systems here due to Shannon s theorem the date rate can be fixed if it is less than channel capacity. (See Lecture 15 Capacity of SG and WM systems ). 6. The more is the number of embedded bits m in the Gaussian case the less is amplitude of embedding that results in turn in problems for practical implementation. 13
Spread-time stegosystems (SG-ST) [4]. b C ( n) = C( n) + ( 1) σπ( n) ith the probability P C ( ) ( ) 1 n = C n ith the probability P (35) Fig 1. SG-ST system ith embedding into pseudorandom samples. (Here = 8, = 38, P 8 38) In order to take a decision about presence or absence of SG under the condition of knon Cn ( ) he (or she) has to perform a testing of to hypothesis: s (36) 14
Optimal hypothesis testing based on maximum likelihood ratio is ( ) λ H ( ) ΛΛ Λ, ΛΛ Λ < λ H, here 1 1 1 ( ) ΛΛ Λ = 1 Suboptimal testing: P( δ H1) P( δ H )!Λ! λ H 1, Λ! <! λ H,% ' here Λ! = 1 & δ (n). ' ( n=1 (37) This method is intuitively reasonable. (38) The efficiency of SG-ST can be estimated (as usually) by to probabilities: P m the probability of SG - ST missing, P fa the probability of SG - ST false detection (false alarm). 15
It as proved in [4 ] that if e let (for simplicity) that folloing loer bound for holds true: P s, η η P Q or P Q here If s P P = P = P, then the (39) σ c η =, s = P (the total number of samples ith embedding) σ as, then P 1 and hence SG - STapproaches to undetectable SG. The probability of bit error after its extraction by legal user (the relation (31))orks. m fa Example. = = = P = = 7 1, η, 1,.45( s 456). Then 15 bits can be embedded ith the bit error probability at most 3 and 1 P.4 16
The use of error correcting codes in SG-ST Then an embedding procedure has to be performed as follos: C ( n ) = C( n ) + ( 1) σπ( n ) ith the probability P j j j C ( n ) = C( n ) ith the probability 1 P l j j b ij b j i = l, here denotes the - th bit the - th codeord of the lenght ij is some integer value. Decoder takes a decision about the embedding of the bij i= argmax ( Cʹ ( n) C( n))( 1) π ( n) 1 i k n= 1 The probability of block error be expressed as P be k d d Pbe ( 1) Q exp R ln, η + η here R= k i -th codeord by making based on ell knon union bound [43 ] can (4) (41) Example. = 1, P.4, η =, = 1, P 1. 7 3 be Then using simplex code (13,1,51) it is possible to embed 44 bits. 17
Optimal detecting of SG-ST Folloing to eq.(37) and to notations (36) e get the logarithmic normalized likelihood ratio as Since is sufficiently large, e can apply the Central Limit Theorem to the sum in (4).Then e get for such choice of the threshold λ, hich provides Pm = Pfa = P the folloing upper bound here, for j =,1 1 1 ΛL ( Λ1 Λ ) = log Pexp δ ( n) (1 P + ) n= 1 ησ W ε # P Q!µ!µ & 1 % $!σ ( ' (!!! $ $ $ *!µ j = E log P exp δ(n) # # " η W σ & + ( 1 P ) *# & " " ε % & % )* %!σ = 1! Var #! log! P $ $ δ(n) # exp # " η W σ & + ( 1 P ) # & " ε % " %! = 1 (! # E *#! log! P $ $ δ(n) # # exp # " η W σ & + ( 1 P ) * # & # " ε % " % )* n=1 " (4) (43) here the random values δ ( n) have the probability distributions given by (36). n=1 n=1 H j + - H j -,- $ & H & = % $ + $ &- & -!µ & & %,- & % 18
Since it is very hard to find analytically the values!µ,!µ 1 and!σ, e estimate them just by the simulation. σ ε η " P = S!m!m 1!σ P = Q!m!m 1 $ #!σ % ' &.1431.161414.16667.4398.41753 1 5.3578.157674.15881.917.41759 4 1 1.7156.15646.157585.5445.41737.1431.161414.1659.4398.41754 5 5.3578.157675.15881.917.41745 1.7156.15646.157583.5445.4176.456.51618.513737.7671.41741 1 5.1131.533.51449.7971.4189 5 1 1.63.49667.49783.718483.41737.456.51618.513854.7671.4177 5 5.1131.4996.577.77769.4186 1.63.49667.497795.718483.41745 19
6 1 7 1 σ ε 1 5 1 η S.1431 5.3578 1.7156.1431 5.3578 1.7156 "!m P =!m 1!σ P = Q!m!m % 1 $ #!σ ' &.16 88.1584 79.1576 86.16 88.1584 79.1576 86.163 93.1585 85.1578 8.164 35.1585 98.1577 97.433 41.314 4.83 97.431 87.314 4.83 97.41835.4186.41548.4175.41711.41461.456 5.135 1 5.41964 5 5.13615 1 7.69844 1 5 5.1131 5.1145 1 5.419 5 5.146 1 7.3173 1 5 1.63 4.97464 1 5.41777 5 4.97587 1 7.836 1 5.456 5.1366 1 5.4181 5 5.135 1 5 7.69844 1 5 5.1131 5.1145 1 5 5.145 1 5 7.3173 1 5.41969 1.63 4.97569 1 5 7.836 1 5.41686 4.97464 1 5
It can be seen that the use of the optimal decision rule does not break undetectability of SG-ST, hence it can be declared as a secure SG system indeed. 1
The aveforms of audio signal after its passing over noisy channel ith signal-to-noise ratio 1 db (a), and the ave form of the same signal after embedding by SG-ST algorithm ith WR=dB (b). (The arros sho the samples ith embedding) We can see that neither visually nor by ears the presence of some embedding cannot be detected. The fact of undetectability of SG-ST by optimal statistical methods has been proved before.