Information Leakage of Correlated Source Coded Sequences over a Channel with an Eavesdropper

Information Leakage of Correlated Source Coded Sequences over a Channel with an Eavesdropper Reevana Balmahoon and Ling Cheng School of Electrical and Information Engineering University of the Witwatersrand Private Bag 3, Wits. 2050, Johannesburg, South Africa Email: reevana.balmahoon@students.wits.ac.za, ling.cheng@wits.ac.za arxiv:40.6264v2 [cs.it] 2 Oct 204 Abstract A new generalised approach for multiple correlated sources over a wiretap network is investigated. A basic model consisting of two correlated sources where each produce a component of the common information is initially investigated. There are several cases that consider wiretapped syndromes on the transmission links and based on these cases a new quantity, the information leakage at the source/s is determined. An interesting feature of the models described in this paper is the information leakage quantification. Shannon s cipher system with eavesdroppers is incorporated into the two correlated sources model to minimize key lengths. These aspects of quantifying information leakage and reducing key lengths using Shannon s cipher system are also considered for a multiple correlated source network approach. A new scheme that incorporates masking using common information combinations to reduce the key lengths is presented and applied to the generalised model for multiple sources. I. INTRODUCTION Keeping information secure has become a major concern with the advancement in technology. In this work, the information theory aspect of security is analyzed, as entropies are used to measure security. The system also incorporates some traditional ideas surrounding cryptography, namely Shannon s cipher system and adversarial attackers in the form of eavesdroppers. In cryptographic systems, there is usually a message in plaintext that needs to be sent to a receiver. In order to secure it, the plaintext is encrypted so as to prevent eavesdroppers from reading its contents. This ciphertext is then transmitted to the receiver. Shannon s cipher system (mentioned by Yamamoto []) incorporates this idea. The definition of Shannon s cipher system has been discussed by Hanawal and Sundaresan [2]. In Yamamoto s [] development on this model, a correlated source approach is introduced. This gives an interesting view of the problem, and is depicted in Figure. Correlated source coding incorporates the lossless compression of two or more correlated data streams. Correlated sources have the ability to decrease the bandwidth required to transmit and receive messages because a syndrome (compressed form of the original message) is sent across the communication links instead of the original message. A compressed message has more information per bit, and therefore has a higher entropy because the transmitted information is more unpredictable. The unpredictability of the compressed message is also beneficial in terms of securing the information. Key Generator Figure. Source W k X, Y W Encoder Wiretapper Decoder Yamamoto s development of the Shannon Cipher System The source sends information for the correlated sources, X and Y along the main transmission channel. A key W k, is produced and used by the encoder when producing the ciphertext. The wiretapper has access to the transmitted codeword, W. The decoded codewords are represented by X and Ŷ. In Yamamoto s scheme the security level was also focused on and found to be K H(XK, Y K W ) (i.e. the joint entropy of X and Y given W, where K is the length of X and Y ) when X and Y have equal importance, which is in accordance with traditional Shannon systems where the security is measured by the equivocation. When one source is more important than the other then the security level is measured by the pair of the individual uncertainties ( K H(XK W ), K H(Y K W )). In practical communication systems links are prone to eavesdropping and as such this work incorporates wiretapped channels, i.e. channels where an eavesdropper is present. There are specific kinds of wiretapped channels that have been developed. The mathematical model for this Wiretap Channel is given by Rouayheb et al. [3], and can be explained as follows: the channel between a transmitter and receiver is error-free and can transmit n symbols Y = (y,..., y n ) from which µ bits can be observed by the eavesdropper and the maximum secure rate can be shown to equal n µ bits. The security aspect of wiretap networks have been looked at in various ways by Cheng et al. [4], and Cai and Yeung [5], emphasising that it is of concern to secure these type of channels. Villard and Piantanida [6] also look at correlated sources and wireap networks: A source sends information to the receiver and an eavesdropper has access to information correlated to the source, which is used as side information. There is a second encoder that sends a compressed version of its own correlation observation of the source privately to the receiver. Here, the authors show that the use of correlation decreases ˆXŶ

2 the required communication rate and increases secrecy. Villard et al. [7] explore this side information concept further where security using side information at the receiver and eavesdropper is investigated. Side information is generally used to assist the decoder to determine the transmitted message. An earlier work involving side information is that by Yang et al. [8]. The concept can be considered to be generalised in that the side information could represent a source. It is an interesting problem when one source is more important and Hayashi and Yamamoto [9] consider it in another scheme, where only X is secure against wiretappers and Y must be transmitted to a legitimate receiver. They develop a security criterion based on the number of correct guesses of a wiretapper to retrieve a message. In an extension of the Shannon cipher system, Yamamoto [0] investigated the secret sharing communication system. In this case, we generalise a model for correlated sources across a channel with an eavesdropper and the security aspect is explored by quantifying the information leakage and reducing the key lengths when incorporating Shannon s cipher system. This paper initially describes a two correlated source model across wiretapped links, which is detailed in Section II. In Section III, the information leakage is investigated and proven for this two correlated source model. The information leakage is quantified to be the equivocation subtracted from the total obtained uncertainty. In Section IV the two correlated sources model is looked at according to Shannon s cipher system. The notation contained in the tables will be clarified in the following sections. The proofs for this Shannon cipher system aspect are detailed in Section V. Section VI details the extension of the two correlated source model where multiple correlated sources in a network scenario is investigated. There are two subsections here; one quantifying information leakage for the Slepian-Wolf scenario and the other incorporating Shannon s cipher system where key lengths are minimized and a masking method to save on keys is presented. Section VII explains how the models detailed in this paper are a generalised model of Yamamoto s [] model, and further offers comparison to other models. The future work for this research is detailed in Section VIII and the paper is concluded in Section IX. II. MODEL The independent, identically distributed (i.i.d.) sources X and Y are mutually correlated random variables, depicted in Figure 2. The alphabet sets for sources X and Y are represented by X and Y respectively. Assume that (X K, Y K ) are encoded into two syndromes (T X and T Y ). We can write T X = (V X, V CX ) and T Y = (V Y, V CY ) where T X and T Y are the syndromes of X and Y. Here, T X and T Y are characterised by (V X, V CX ) and (V Y, V CY ) respectively. The Venn diagram in Figure 3 easily illustrates this idea where it is shown that V X and V Y represent the private information of sources X and Y respectively and V CX and V CY represent the common information between X K and Y K generated by X K and Y K respectively. The correlated sources X and Y transmit messages (in the form of syndromes) to the receiver along wiretapped links. Figure 2. Figure 3. X K Encoder T X Decoder ˆX K, Ŷ K Y K Encoder T Y Correlated source coding for two sources X V X V CX V CY V Y T X = (V X, V CX ) T Y = (V Y, V CY ) The relation between private and common information The decoder determines X and Y only after receiving all of T X and T Y. The common information between the sources are transmitted through the portions V CX and V CY. In order to decode a transmitted message, a source s private information and both common information portions are necessary. This aids in security as it is not possible to determine, for example X by wiretapping all the contents transmitted along X s channel only. This is different to Yamamoto s [] model as here the common information consists of two portions. The aim is to keep the system as secure as possible and these following sections show how it is achieved by this new model. We assume that the function F is a one-to-one process with high probability, which means based on T X and T Y we can retrieve X K and Y K with minimal error. Furthermore, it reaches the Slepian-Wolf bound, H(T X, T Y ) = H(X K, Y K ). Here, we note that the lengths of T X and T Y are not fixed, as it depends on the encoding process and nature of the Slepian- Wolf codes. The process is therefore not ideally one-to-one and reversible and is another difference between our model and Yamamoto s [] model. The code described in this section satisfies the following inequalities for δ > 0 and sufficiently large K. P r{x K G(V X, V CX, V CY )} δ () P r{y K G(V Y, V CX, V CY )} δ (2) H(V X, V CX, V CY ) H(X K ) + δ (3) H(V Y, V CX, V CY ) H(Y K ) + δ (4) Y

3 H(V X, V Y, V CX, V CY ) H(X K, Y K ) + δ (5) H(Y X) ɛ 0 K H(W Y ) K log M Y H(Y X) + ɛ 0 (2) H(X K V X, V Y ) H(V CX ) + H(V CY ) δ (6) H(X K V CX, V CY ) H(V X ) + H(V CY ) δ (7) H(X K V CX, V CY, V Y ) H(V X ) + H(V CY ) δ (8) H(V CX ) + H(V X ) δ H(X K V CY, V Y ) H(X) H(V CY ) + δ (9) I(X; Y ) ɛ 0 K (H(W CX) + H(W CY )) K (log M CX + log M CY ) I(X; Y ) + ɛ 0 (3) K H(XK W Y ) H(X) ɛ 0 (4) K H(Y K W X ) H(Y ) ɛ 0 (5) We can see that () - (3) mean where G is a function to define the decoding process at the receiver. It can intuitively be seen from (3) and (4) that X and Y are recovered from the corresponding private information and the common information produced by X K and Y K. Equations (3), (4) and (5) show that the private information and common information produced by each source should contain no redundancy. It is also seen from (7) and (8) that V Y is independent of X K asymptotically. Here, V X, V Y, V CX and V CY are disjoint, which ensures that there is no redundant information sent to the decoder. To recover X the following components are necessary: V X, V CX and V CY. This comes from the property that X K cannot be derived from V X and V CX only and part of the common information between X K and Y K is produced by Y K. Yamamoto [] proved that a common information between X K and Y K is represented by the mutual information I(X; Y ). Yamamoto [] also defined two kinds of common information. The first common information is defined as the rate of the attainable minimum core V C (i.e. V CX, V CY in this model) by removing each private information, which is independent of the other information, from (X K, Y K ) as much as possible. The second common information is defined as the rate of the attainable maximum core V C such that if we lose V C then the uncertainty of X and Y becomes H(V C ). Here, we consider the common information that V CX and V CY represent. We begin demonstrating the relationship between the common information portions by constructing the prototype code (W X, W Y, W CX, W CY ) as per Lemma. Lemma : For any ɛ 0 0 and sufficiently large K, there exits a code W X = F X (X K ), W Y = F Y (Y K ), W CX = F CX (X K ), W CY = F CY (Y K ), XK, Ŷ K = G(W X, W Y, W CX, W CY ), where W X I MX, W Y I MY, W CX I MCX, W CY I MCY for I Mα, which is defined as {0,,..., M α }, that satisfies, P r{ X K, Ŷ K X K, Y K } ɛ (0) H(X Y ) ɛ 0 K H(W X) K log M shown by Yamamoto []. X H(X Y ) + ɛ 0 () H(X, Y ) 3ɛ 0 K (H(W X) + H(W Y ) + H(W CX ) + H(W CY )) H(X, Y ) + 3ɛ 0 (6) Hence from (0), (6) and the ordinary source coding theorem, (W X, W Y, W CX, W CY ) have no redundancy for sufficiently small ɛ 0 0. It can also be seen that W X and W Y are independent of Y K and X K respectively. Proof of Lemma : As seen by Slepian and Wolf, mentioned by Yamamoto [] there exist M X codes for the P Y X (y x) DMC (discrete memoryless channel) and M Y codes for the P X Y (x y) DMC. The codeword sets exist as Ci X and Cj Y, where CX i is a subset of the typical sequence of X K and Cj Y is a subset of the typical sequence of Y K. The encoding functions are similar, but we have created one decoding function as there is one decoder at the receiver: f Xi : I MCX C X i (7) f Y j : I MCY C Y j (8) g : X K, Y K I MCX I MCY (9) The relations for M X, M Y and the common information remain the same as per Yamamoto s and will therefore not be proven here. In this scheme, we use the average (V CX, V X, V CY, V Y ) transmitted for many codewords from X and Y. Thus, at any time either V CX or V CY is transmitted. Over time, the split between which common information portion is transmitted is determined and the protocol is prearranged accordingly. Therefore all the common information is either transmitted as l or m, and as such Yamamoto s encoding and decoding method may be used. As per Yamamoto s method the code does exist and that W X and W Y are independent of Y and X respectively, as

4 The common information is important in this model as the sum of V CX and V CY represent a common information between the sources. The following theorem holds for this common information: Theorem : K [H(V CX) + H(V CY )] = I(X; Y ) (20) where V CX is the common portion between X and Y produced by X K and V CY is the common portion between X and Y produced by Y K. It is noted that the (20) holds asymptotically, and does not hold with equality when K is finite. Here, we show the approximation when K is infinitely large. The private portions for X K and Y K are represented as V X and V Y respectively. As explained in Yamamoto s [] Theorem, two types of common information exist (the first is represented by I(X; Y ) and the second by min(h(x K ), H(Y K )). We will develop part of this idea to show that the sum of the common information portions produced by X K and Y K in this new model is represented by the mutual information between the sources. Proof of Theorem : The first part is to prove that H(V CX ) + H(V CY ) I(X; Y ), and is done as follows. We weaken the conditions () and (2) to Pr {X K, Y K G XY (V X, V Y, V CX, V CY }) δ (2) For any (V X,V Y, V CX, V CY ) C(3ɛ 0 ) (which can be seen from (6)), we have from (2) and the ordinary source coding theorem that H(X K, Y K ) δ K H(V X, V Y, V CX, V CY ) K [H(V X) + H(V Y ) + H(V CX ) + H(V CY )] (22) where δ 0 as δ 0. From Lemma, K H(V Y X K ) K H(V Y ) δ (23) K H(V X Y K ) K H(V X) δ (24) From (22) - (24), K [H(V CX) + H(V CY )] H(X, Y ) K H(V X) K H(V Y ) δ H(X, Y ) K H(V X Y ) On the other hand, we can see that K H(V Y X) δ 2δ(25) K H(XK, V Y ) H(X, Y ) + δ (26) This implies that K H(V Y X K ) H(Y X) + δ (27) and K H(V X Y K ) H(X Y ) + δ (28) From (25), (27) and (28) we get K [H(V CX) + H(V CY )] H(X, Y ) H(X Y ) H(Y X) δ 4δ = I(X; Y ) δ 4δ (29) It is possible to see from (3) that H(V CX ) + H(V CY ) I(X; Y ). From this result, (9) and (29), and as δ 0 and δ 0 it can be seen that K [H(V CX + H(V CY )] = I(X; Y ) (30) This model can cater for a scenario where a particular source, say X needs to be more secure than Y (possibly because of eavesdropping on the X channel). In such a case, the K H(V CX) term in (29) needs to be as high as possible. When this uncertainty is increased then the security of X is increased. Another security measure that this model incorporates is that X cannot be determined from wiretapping only X s link. III. INFORMATION LEAKAGE In order to determine the security of the system, a measure for the amount of information leaked has been developed. This is a new notation and quantification, which emphasizes the novelty of this work. The obtained information and total uncertainty are used to determine the leaked information. Information leakage is indicated by L P Q. Here P indicates the source/s for which information leakage is being quantified, P = {S,..., S n } where n is the number of sources (in this case, n = 2). Further, Q indicates the syndrome portion that has been wiretapped, Q = {V,..., V m } where m is the number of codewords (in this case, m = 4). The information leakage bounds are as follows: L XK V X,V Y H(X K ) H(V CX ) H(V CY ) + δ (3) L XK V CX,V CY H(X K ) H(V X ) H(V CY ) + δ (32) L XK V CX,V CY,V Y H(X K ) H(V X ) H(V CY ) + δ (33) H(V CY ) δ L XK V Y,V CY H(X K ) H(V CX ) H(V X ) + δ (34) Here, V Y is private information of source Y K and is independent of X K and therefore does not leak any information about X K, shown in (32) and (33). Equation (34) gives an indication of the minimum and maximum amount of leaked information for the interesting case where a syndrome has been wiretapped and its information leakage on the alternate source is quantified. The outstanding common information component

5 is the maximum information that can be leaked. For this case, the common information V CX and V CY can thus consist of added protection to reduce the amount of information leaked. These bounds developed in (3) - (34) are proven in the next section. The proofs for the above mentioned information leakage inequalities are now detailed. First, the inequalities in (6) - (9) will be proven, so as to prove that the information leakage equations hold. Lemma 2: The code (V X, V CX, V CY, V Y ) defined at the beginning of Section I, describing the model and () - (5) satisfy (6) - (9). Then the information leakage bounds are given by (3) - (34). Proof for (6): K H(XK V X, V Y ) = K [H(XK, V X, V Y ) H(V X, V Y )] = K [H(XK, V Y ) H(V X, V Y )] (35) = K [H(XK V Y ) + I(X K ; V Y ) + H(V Y X K )] K [H(V X V Y ) + I(V X ; V Y ) + H(V Y V X )] = K [H(XK V Y ) + H(V Y X K ) H(V X V Y ) H(V Y V X )] = K [H(XK ) + H(V Y ) H(V X ) H(V Y )] (36) = K [H(XK ) H(V X )] K [H(V X) + H(V CX ) + H(V CY ) H(V X )] δ = K [H(V CX) + H(V CY )] δ (37) where (35) holds because V X is a function of X and (36) holds because X is independent of V Y asymptotically and V X is independent of V Y asymptotically. For the proofs of (7) and (8), the following simplification for H(X V CY ) is used: Proof for (7): K H(XK V CX, V CY ) = K [H(XK, V CX, V CY ) H(V CX, V CY )] = K [H(XK, V CY ) H(V CX, V CY )] (40) = K [H(XK ) H(V CY ) + I(X; V CY ) + H(V CY X K )] + δ K [H(V CX V CY ) + I(V CX ; V CY ) + H(V CY V CX )] = K [H(XK ) H(V CY ) + H(V CY ) H(V CX ) H(V CY )] + δ (4) = K [H(XK ) H(V CY ) H(V CX )] + δ K [H(V X) + H(V CX ) + H(V CY ) H(V CY ) H(V CX )] δ = K H(V X) + δ δ (42) where (40) holds because V CX is a function of X K and (4) holds because X is independent of V CY asymptotically and V CX is independent of V CY asymptotically. The proof for H(X V CX, V CY, V Y ) is similar to that for H(X V CX, V CY ), because V Y is independent of X. Proof for (8): K H(XK V CX, V CY, V Y ) = K H(XK V CX, V CY ) (43) = K [H(XK, V CX, V CY ) H(V CX, V CY )] = K [H(XK, V CY ) H(V CX, V CY )] (44) = K [H(XK ) H(V CY ) + I(X; V CY ) + H(V CY X K )] + δ K [H(V CX V CY ) + I(V CX ; V CY ) + H(V CY V CX )] = K [H(XK ) H(V CY ) + H(V CY ) H(V CX ) H(V CY )] + δ (45) = H(X K V CY ) = H(X K, Y K ) H(V CY ) K [H(XK ) H(V CY ) H(V CX )] + δ = H(X K ) + H(V CY ) I(X; V CY ) H(V CY ) K [H(V X) + H(V CX ) + H(V CY ) H(V CY ) = H(X K ) + H(V CY ) H(V CY ) H(V CY ) H(V CX )] δ + δ + δ (38) = = H(X K K H(V X) δ + δ (46) ) H(V CY ) = δ (39) where (43) holds because V Y and X K are independent, (44) holds because V CX is a function of X K and (45) holds where I(X; V CY ) approximately equal to H(V CY ) in (38) because X K is independent of V CY asymptotically and V CX can be seen intuitively from the Venn diagram in Figure is independent of V CY asymptotically. 3. Since it is an approximation, δ, which is smaller than For the proof of (9), we look at the following probabilities: δ in the proofs below has been added to cater for the tolerance. Pr{V X, V CX G(T X )} δ (47)

6 Pr{V Y, V CY G(T Y )} δ (48) K H(XK T Y ) K H(XK, V CY, V Y )] + δ (49) = K [H(XK, V CY, V Y ) H(V CY, V Y )] + δ = K [H(XK, V Y ) H(V CY, V Y )] + δ (50) = K [H(XK V Y ) + I(X K ; V Y ) + H(V Y X K )] K [H(V CY V Y ) + I(V CY ; V Y ) + H(V Y V CY )] + δ = K [H(XK ) + H(V Y ) H(V CY ) H(V Y )] + δ (5) = K [H(XK ) H(V CY )] + δ (52) where (49) holds from (48), (50) holds because V CY and V Y are asymptotically independent. Furthermore, (5) holds because V CY and V Y are asymptotically independent and X K and V Y are asymptotically independent. Following a similar proof to those done above in this section, another bound for H(X K V CY, V Y ) can be found as follows: K H(XK V CY, V Y ) = K [H(XK, V CY, V Y ) H(V CY, V Y )] = K [H(XK, V Y ) H(V CY, V Y )] (53) = K [H(XK V Y ) + I(X K ; V Y ) + H(V Y X)] K [H(V CY V Y ) + I(V CY ; V Y ) + H(V Y V CY )] = K [H(XK ) + H(V Y ) H(V CY ) H(V Y )] (54) = K [H(XK ) H(V CY )] K [H(V X) + H(V CX ) + H(V CY ) H(V CY )] δ which proves (32). L XK V CX,V CY,V Y = H(X K ) H(X K V CX, V CY, V Y ) H(X K ) H(V X ) + δ (58) which proves (33). The two bounds for H(V CY, V Y ) are given by (52) and (55). From (52): L XK V Y,V CY H(X K ) [H(X) H(V CY ) + δ] H(V CY ) δ (59) and from (55): L XK V Y,V CY H(X K ) (H(V X ) + H(V CX ) δ) H(X K ) H(V X ) H(V CX ) + δ (60) Combining these results from (59) and (60) gives (34). IV. SHANNON S CIPHER SYSTEM Here, we discuss Shannon s cipher system for two independent correlated sources (depicted in Figure 4). The two source outputs are i.i.d random variables X and Y, taking on values in the finite sets X and Y. Both the transmitter and receiver have access to the key, a random variable, independent of X K and Y K and taking values in I Mk = {0,, 2,..., M k }. The sources X K and Y K compute the ciphertexts X and Y, which are the result of specific encryption functions on the plaintext from X and Y respectively. The encryption functions are invertible, thus knowing X and the key, X K can be retrieved. The mutual information between the plaintext and ciphertext should be small so that the wiretapper cannot gain much information about the plaintext. For perfect secrecy, this mutual information should be zero, then the length of the key should be at least the length of the plaintext. X K Y K = K [H(V X) + H(V CX )] δ (55) where (53) and (54) hold for the same reason as (50) and (5) respectively. Since we consider the information leakage as the total information obtained subtracted from the total uncertainty, the following hold for the four cases considered in this section: L XK V X,V Y = H(X K ) H(X K V X, V Y ) H(X K ) H(V CX ) H(V CY ) + δ (56) which proves (3). Figure 4. k k Encoder Encoder X Y k Decoder ˆX K, Ŷ K Shannon cipher system for two correlated sources L XK V CX,V CY = H(X K ) H(X K V CX, V CY ) H(X K ) H(V X ) + δ (57) respec- The encoder functions for X and Y, (E X and E Y tively) are given as:

7 E X : X K I MkX I M X = {0,,..., M X } I M CX = {0,,..., M CX }(6) K H(XK, Y K W ) h XY + ɛ (77) E Y : Y K I MkY I M Y = {0,,..., M Y } The decoder is defined as: I M CY = {0,,..., M CY }(62) D XY : (I M X, I M Y, I M CX, I M CY ) I MkX, I MkY The encoder and decoder mappings are below: or X K Y K (63) W = F EX (X K, W kx ) (64) W 2 = F EY (Y K, W ky ) (65) X K = F DX (W, W 2, W kx ) (66) Ŷ K = F DY (W, W 2, W ky ) (67) ( X K, Ŷ K ) = F DXY (W, W 2, W kx, W ky ) (68) The following conditions should be satisfied for cases - 4: K log M X R X + ɛ (69) K log M Y R Y + ɛ (70) K log M kx R kx + ɛ (7) K log M ky R ky + ɛ (72) Pr{ X K X K } ɛ (73) Pr{Ŷ K Y K } ɛ (74) K H(XK W ) h X + ɛ (75) K H(Y K W 2 ) h Y + ɛ (76) K H(XK, Y K W 2 ) h XY + ɛ (78) where R X is the the rate of source X s channel and R Y is the the rate of source Y s channel. Here, R kx is the rate of the key channel at X K and R ky is the rate of the key channel at Y K. The security levels, which are measured by the total and individual uncertainties are h XY and (h X, h Y ) respectively. The cases - 5 are: Case : When T X and T Y are leaked and both X K and Y K need to be kept secret. Case 2: When T X and T Y are leaked and X K needs to be kept secret. Case 3: When T X is leaked and both X K and Y K need to be kept secret. Case 4: When T X is leaked and Y K needs to be kept secret. Case 5: When T X is leaked and X K needs to be kept secret. where T X is the syndrome produced by X, containing V CX and V X and T Y is the syndrome produced by Y, containing V CY and V X. The admissible rate region for each case is defined as follows: Definition a: (R X, R Y, R kx, R ky, h XY ) is admissible for case if there exists a code (F EX, F DXY ) and (F EY, F DXY ) such that (69) - (74) and (78) hold for any ɛ 0 and sufficiently large K. Definition b: (R X, R Y, R kx, R ky, h X ) is admissible for case 2 if there exists a code (F EX, F DXY ) such that (69) - (75) hold for any ɛ 0 and sufficiently large K. Definition c: (R X, R Y, R kx, R ky, h X, h Y ) is admissible for case 3 if there exists a code (F EX, F DXY ) and (F EY, F DXY ) such that (69) - (74) and (76), (78) hold for any ɛ 0 and sufficiently large K. Definition d: (R X, R Y, R kx, R ky, h Y ) is admissible for case 4 if there exists a code (F EX, F DXY ) such that (69) - (74) and (76) hold for any ɛ 0 and sufficiently large K. Definition e: (R X, R Y, R kx, R ky, h X ) is admissible for case 5 if there exists a code (F EX, F DXY ) such that (69) - (74) and (75) hold for any ɛ 0 and sufficiently large K. Definition 2: The admissible rate regions of R j and of R k are defined as: R (h XY ) = {(R X, R Y, R kx, R ky ) : (R X, R Y, R kx, R ky, h XY ) is admissible for case } (79) R 2 (h X ) = {(R X, R Y, R kx, R ky ) : (R X, R Y, R kx, R ky, h X ) is admissible for case 2} (80)

8 R 3 (h X, h Y ) = {(R X, R Y, R kx, R ky ) : (R X, R Y, R kx, R ky, h X, h Y ) is admissible for case 3} (8) R 4 (h Y ) = {(R X, R Y, R kx, R ky ) : (R X, R Y, R kx, R ky, h Y ) is admissible for case 4} (82) R 5 (h X ) = {(R X, R Y, R kx, R ky ) : (R X, R Y, R kx, R ky, h X ) is admissible for case 5} (83) Theorems for these regions have been developed: Theorem 2: For 0 h XY H(X, Y ), R (h XY ) = {(R X, R Y, R kx, R ky ) : R X H(X Y ), R Y H(Y X), R X + R Y H(X, Y ) R kx h XY and R ky h XY } (84) Theorem 3: For 0 h X H(X), R 2 (h X ) = {(R X, R Y, R kx, R ky ) : R X H(X Y ), R Y H(Y X), R X + R Y H(X, Y ) R kx h X and R ky h Y } (85) Theorem 4: For 0 h X H(X) and 0 h Y H(Y ), R 3 (h X, h Y ) = {(R X, R Y, R kx, R ky ) : R X H(X Y ), R Y H(Y X), R X + R Y H(X, Y ) R kx h X and R ky h Y } (86) Theorem 5: For 0 h X H(X), R 5 (h X, h Y ) = {(R X, R Y, R kx, R ky ) : R X H(X Y ), R Y H(Y X), R X + R Y H(X, Y ) R kx h X and R ky 0} (87) When h X = 0 then case 5 can be reduced to that depicted in (86). Hence, Corollary follows: Corollary : R 4 (h Y ) = R 3 (0, h Y ) The security levels, which are measured by the total and individual uncertainties h XY and (h X, h Y ) respectively give an indication of the level of uncertainty in knowing certain information. When the uncertainty increases then less information is known to an eavesdropper and there is a higher level of security. V. PROOF OF THEOREMS 2-5 This section initially proves the direct parts of Theorems 2-5 and thereafter the converse parts. A. Direct parts All the channel rates in the theorems above are in accordance with Slepian-Wolf s theorem, hence there is no need to prove them. We construct a code based on the prototype code (W X, W Y, W CX, W CY ) in Lemma. In order to include a key in the prototype code, W X is divided into two parts as per the method used by Yamamoto []: W X = W X mod M X I MX = {0,, 2,..., M X } (88) W X2 = W X W X M X I MX2 = {0,, 2,..., M X2 } (89) where M X is a given integer and M X2 is the ceiling of M X /M X. The M X /M X is considered an integer for simplicity, because the difference between the ceiling value and the actual value can be ignored when K is sufficiently large. In the same way, W Y is divided: W Y = W Y mod M Y I MY = {0,, 2,..., M Y } (90) W Y 2 = W Y W Y M Y I MY 2 = {0,, 2,..., M Y 2 } (9) The common information components W CX and W CY are already portions and are not divided further. It can be shown that when some of the codewords are wiretapped the uncertainties of X K and Y K are bounded as follows: K H(XK W X2, W Y ) I(X; Y ) + K log M X ɛ 0 (92) K H(Y K W X, W Y 2 ) I(X; Y ) + K log M Y ɛ 0 (93) K H(XK W X, W Y 2 ) I(X; Y ) ɛ 0 (94) K H(XK W X, W Y, W CY ) K log M CX ɛ 0 (95) K H(Y K W X, W Y, W CY ) K log M CX ɛ 0 (96) K H(XK W Y, W CY ) H(X Y ) + K log M CX ɛ 0 (97) K H(Y K W Y, W CY ) K log M CX ɛ 0 (98) where ɛ 0 0 as ɛ 0 0. The proofs for (92) - (98) are the same as per Yamamoto s [] proof in Lemma A. The

9 difference is that W CX, W CY, M CX and M CY are described as W C, W C2, M C and M C2 respectively by Yamamoto. Here, we consider that W CX and W CY are represented by Yamamoto s W C and W C2 respectively. In addition there are some more inequalities considered here: K H(Y K W X, W CX, W CY, W Y 2 ) K log M Y ɛ 0 (99) The codewords W X and W Y and their keys W kx and W ky are now defined: W X = (W X W kcy, W X2 W kcy, W CX W kcy )(06) W Y = (W Y W kcy, W Y 2 W kcy, W CY ) (07) K H(Y K W X, W CX, W CY ) K log M Y + K log M Y 2 ɛ 0(00) K H(XK W X2, W CY ) K log M X K H(Y K W X2, W CY ) K log M Y + K log M CX ɛ 0 (0) + K log M Y 2 + K log M CX ɛ 0 (02) The inequalities (99) and (00) can be proved in the same way as per Yamamoto s [] Lemma A2, and (0) and (02) can be proved in the same way as per Yamamoto s [] Lemma A. For each proof we consider cases where a key already exists for either V CX or V CY and the encrypted common information portion is then used to mask the other portions (either V CX or V CY and the private information portions). There are two cases considered for each; firstly, when the common information portion entropy is greater than the entropy of the portion that needs to be masked, and secondly when the common information portion entropy is less than the entropy of the portion to be masked. For the latter case, a smaller key will need to be added so as to cover the portion entirely. This has the effect of reducing the required key length, which is explained in greater detail in Section VII. Proof of Theorem 2: Suppose that (R X, R Y, R KX, R KY ) R for h XY H(X, Y ). Without loss of generality, we assume that h X h Y. Then, from (84) R X H(X K Y K ) R Y H(Y K X K ) R X + R Y H(X K, Y K ) (03) R kx h XY, R ky h XY (04) Assuming a key exists for V CY. For the first case, consider the following: H(V CY ) H(V X ), H(V CY ) H(V Y ) and H(V CY ) H(V CX ). M CY = 2 Kh XY (05) W ky = (W kcy ) (08) where W α I Mα = {0,,..., M α }. The wiretapper will not know W X, W X2 and W CX from W X and W Y, W Y 2 and W CY from W Y as these are protected by the key (W kcy. In this case, R X, R Y, R kx and R ky satisfy from () - (3) and (03) - (05), that K log M X + K log M Y = K (log M X + log M X2 + log M CX ) + K (log M Y + log M Y 2 + log M CY ) H(X Y ) + H(Y X) + I(X; Y ) + ɛ 0 = H(X, Y ) R X + R Y (09) K log M kx = K log M CY = h XY (0) where (0) comes from (05). R kx () K log M ky = K log M CY = h XY (2) where (2) comes from (05). The security levels thus result: R ky (3) K H(XK W X, W Y ) = K H(X W X W kcy, W X2 W kcy, W CX W kcy, W Y W kcy, W Y 2 W kcy, W CY ) = H(X K ) (4) h X ɛ 0 (5) where (4) holds because W X, W X2, W CX, W Y, W Y 2 are covered by key W CY and W CY is covered by an existing random number key. Equations (0) - (6) imply that W X, W X2, W Y and W Y 2 have almost no redundancy and they are mutually independent.

0 Similarly, K H(Y K W X, W Y ) h Y ɛ 0 (6) Therefore (R X, R Y, R kx, R ky, h XY, h XY ) is admissible from (09) - (6). Next the case where: H(V CY ) < H(V X ), H(V CY ) < H(V Y ) and H(V CY ) < H(V CX ) is considered. Here, there are shorter length keys used in addition to the key provided by W CY in order to make the key lengths required by the individual portions. For example the key W k comprises W kcy and a short key W, which together provide the length of W X. The codewords W X and W Y and their keys W kx and W ky are now defined: W X = (W X W k, W X2 W k2, W CX W k3 ) (7) W Y = (W Y W k4, W Y 2 W k5, W CY ) (8) W kx = (W k, W k2, W k3 ) (9) W ky = (W k4, W k5 ) (20) where W α I Mα = {0,,..., M α }. The wiretapper will not know W X, W X2 and W CX from W X and W Y, W Y 2 and W CY from W Y as these are protected by the key (W kcy. In this case, R X, R Y, R kx and R ky satisfy that K log M X + K log M Y = K (log M X + log M X2 + log M CX ) + K (log M Y + log M Y 2 + log M CY ) H(X Y ) + H(Y X) + I(X; Y ) + ɛ 0 = H(X, Y ) R X + R Y (2) K log M kx = K [log M k + log M k2 + log M k3 ] = log M kcy + log M + log M kcy + log M 2 + log M kcy + log M 3 = 3 log M kcy + log M + log M 2 + log M 3 3h XY ɛ 0 (22) where (22) results from (05). h XY (23) K log M kx = K [log M k3 + log M k4 + log M kcy ] = log M kcy + log M 3 + log M kcy + log M 4 + log M kcy = 3 log M kcy + log M 3 + log M 4 (24) 3h XY ɛ 0 (25) h XY (26) where (25) results from (05). The security levels thus result: K H(XK W X, W Y ) = K H(X W X W k, W X2 W k2, W CX W k3, W Y W k4, W Y 2 W k5, W CY ) = H(X K ) (27) h X ɛ 0 (28) where (4) holds because W X, W X2, W CX, W Y, W Y 2 are covered by key W CY and some shorter length key and W CY is covered by an existing random number key. Similarly, K H(Y K W X, W Y ) h Y ɛ 0 (29) Therefore (R X, R Y, R kx, R ky, h XY, h XY ) is admissible from (2) - (29). Theorem 3-5 are proven in the same way with varying codewords and keys. The proofs follow: Theorem 3 proof: The consideration for the security levels is that h Y h X because Y contains the key the is used for masking. Suppose that (R X, R Y, R KX, R KY ) R 2. From (85) R X H(X K Y K ) R Y H(Y K X K ) R X + R Y H(X K, Y K ) (30) R kx h X, R ky h Y (3) Assuming a key exists for V CY. For the first case, consider the following: H(V CY ) H(V X ), H(V CY ) H(V Y ) and H(V CY ) H(V CX ). M CY = 2 Kh Y (32) The codewords W X and W Y and their keys W kx and W ky are now defined: W X = (W X W kcy, W X2 W kcy, W CX W kcy )(33) W Y = (W Y W kcy, W Y 2 W kcy, W CY ) (34)

W ky = (W kcy ) (35) where W α I Mα = {0,,..., M α }. The wiretapper will not know W X, W X2 and W CX from W X and W Y, W Y 2 and W CY from W Y as these are protected by the key (W kcy. In this case, R X, R Y, R kx and R ky satisfy from () - (3) and (30) - (32), that K log M X + K log M Y = K (log M X + log M X2 + log M CX ) + K (log M Y + log M Y 2 + log M CY ) H(X Y ) + H(Y X) + I(X; Y ) + ɛ 0 = H(X, Y ) R X + R Y (36) K log M kx = K log M CY = h Y (37) h X ɛ 0 (38) R kx (39) where (37) comes from (32) and (38) comes form the consideration stated at the beginning of this proof. K log M ky = K log M CY = h XY (40) where (40) comes from (32). The security levels thus result: R ky (4) K H(XK W X, W Y ) = K H(X W X W kcy, W X2 W kcy, W CX W kcy, W Y W kcy, W Y 2 W kcy, W CY ) = H(X K ) (42) h X ɛ 0 (43) where (67) holds because W X, W X2, W CX, W Y, W Y 2 are covered by key W CY and W CY is covered by an existing random number key. Equations (0) - (6) imply that W X, W X2, W Y and W Y 2 have almost no redundancy and they are mutually independent. Similarly, K H(Y K W X, W Y ) h Y ɛ 0 (44) Therefore (R X, R Y, R kx, R ky, h XY, h XY ) is admissible from (09) - (6). Next the case where: H(V CY ) < H(V X ), H(V CY ) < H(V Y ) and H(V CY ) < H(V CX ) is considered. Here, there are shorter length keys used in addition to the key provided by W CY in order to make the key lengths required by the individual portions. For example the key W k comprises W kcy and a short key W, which together provide the length of W X. The codewords W X and W Y and their keys W kx and W ky are now defined: W X = (W X W k, W X2 W k2, W CX W k3 ) (45) W Y = (W Y W k4, W Y 2 W k5, W CY ) (46) W kx = (W k, W k2, W k3 ) (47) W ky = (W k4, W k5 ) (48) where W α I Mα = {0,,..., M α }. The wiretapper will not know W X, W X2 and W CX from W X and W Y, W Y 2 and W CY from W Y as these are protected by the key (W kcy. In this case, R X, R Y, R kx and R ky satisfy that K log M X + K log M Y = K (log M X + log M X2 + log M CX ) + K (log M Y + log M Y 2 + log M CY ) H(X Y ) + H(Y X) + I(X; Y ) + ɛ 0 = H(X, Y ) R X + R Y (49) K log M kx = K [log M k + log M k2 + log M k3 ] = log M kcy + log M + log M kcy + log M 2 + log M kcy + log M 3 = 3 log M kcy + log M + log M 2 + log M 3 3h Y ɛ 0 (50) 3h X ɛ 0 h X (5) where (50) results from (32) and the result is from the consideration at the beginning of this proof. K log M ky = K [log M k3 + log M k4 + log M kcy ] = log M kcy + log M 3 + log M kcy + log M 4 + log M kcy = 3 log M kcy + log M 3 + log M 4 3h Y ɛ 0 (52) h Y (53)

2 where (52) results from (32). The security levels thus result: K H(XK W X, W Y ) = K H(X W X W k, W X2 W k2, W CX W k3, W Y W k4, W Y 2 W k5, W CY ) = H(X K ) (54) h X ɛ 0 (55) where (54) holds because W X, W X2, W CX, W Y, W Y 2 are covered by key W CY and some shorter length key and W CY is covered by an existing random number key. Similarly, K H(Y K W X, W Y ) h Y ɛ 0 (56) Therefore (R X, R Y, R kx, R ky, h XY, h XY ) is admissible from (2) - (29). Proof of Theorem 4: Again, the consideration for the security levels is that h Y h X because Y contains the key the is used for masking. Suppose that (R X, R Y, R KX, R KY ) R 3. From (85) R X H(X K Y K ) R Y H(Y K X K ) R X + R Y H(X K, Y K ) (57) R kx h X, R ky h Y (58) Assuming a key exists for V CY. For the first case, consider the following: H(V CY ) H(V X ), H(V CY ) H(V Y ) and H(V CY ) H(V CX ). M CY = 2 Kh Y (59) In the same way as theorem 2 and 3, the codewords W X and W Y and their keys W kx and W ky are now defined: W X = (W X W kcy, W X2 W kcy, W CX W kcy )(60) W Y = (W Y W kcy, W Y 2 W kcy, W CY ) (6) W ky = (W kcy ) (62) where W α I Mα = {0,,..., M α }. The wiretapper will not know W X, W X2 and W CX from W X and W Y, W Y 2 and W CY from W Y as these are protected by the key (W kcy. K log M X + K log M Y = K (log M X + log M X2 + log M CX ) + K (log M Y + log M Y 2 + log M CY ) H(X Y ) + H(Y X) + I(X; Y ) + ɛ 0 = H(X, Y ) R X + R Y (63) K log M kx = K log M CY = h Y (64) h X ɛ 0 (65) R kx (66) where (64) comes from (59) and (65) comes form the consideration stated at the beginning of this proof. K log M ky = K log M CY = h XY (67) where (67) comes from (59). The security levels thus result: R ky (68) K H(XK W X, W Y ) = K H(X W X W kcy, (69) W X2 W kcy, W CX W kcy, W Y W kcy, W Y 2 W kcy, W CY ) = H(X K ) (70) h X ɛ 0 (7) where (70) holds because W X, W X2, W CX, W Y, W Y 2 are covered by key W CY and W CY is covered by an existing random number key. Equations (0) - (6) imply that W X, W X2, W Y and W Y 2 have almost no redundancy and they are mutually independent. Similarly, K H(Y K W X, W Y ) h Y ɛ 0 (72) Therefore (R X, R Y, R kx, R ky, h XY, h XY ) is admissible from (09) - (6). Next the case where: H(V CY ) < H(V X ), H(V CY ) < H(V Y ) and H(V CY ) < H(V CX ) is considered. Here, there are shorter length keys used in addition to the key provided by W CY in order to make the key lengths required by the individual portions. For example the key W k comprises W kcy and a short key W, which together provide the length of W X. The codewords W X and W Y and their keys W kx and W ky are now defined: W X = (W X W k, W X2 W k2, W CX W k3 ) (73)

3 W Y = (W Y W k4, W Y 2 W k5, W CY ) (74) W kx = (W k, W k2, W k3 ) (75) W ky = (W k4, W k5 ) (76) where W α I Mα = {0,,..., M α }. The wiretapper will not know W X, W X2 and W CX from W X and W Y, W Y 2 and W CY from W Y as these are protected by the key (W kcy. In this case, R X, R Y, R kx and R ky satisfy that K log M X + K log M Y = K (log M X + log M X2 + log M CX ) + K (log M Y + log M Y 2 + log M CY ) H(X Y ) + H(Y X) + I(X; Y ) + ɛ 0 = H(X, Y ) R X + R Y (77) K log M kx = K [log M k + log M k2 + log M k3 ] = log M kcy + log M + log M kcy where (83) holds because W X, W X2, W CX, W Y, W Y 2 are covered by key W CY and some shorter length key and W CY is covered by an existing random number key. Similarly, K H(Y K W X, W Y ) h Y ɛ 0 (85) Therefore (R X, R Y, R kx, R ky, h XY, h XY ) is admissible from (77) - (85). The region indicated for R is derived from this region for R, when h X = 0. Proof of Theorem 5: As before, V CY may be used as a key, however here we use V CX as the key in this proof to show some variation. Now the consideration for the security levels is that h X h Y because X contains the key that is used for masking. Suppose that (R X, R Y, R KX, R KY ) R 5. From (85) R X H(X K Y K ) R Y H(Y K X K ) R X + R Y H(X K, Y K ) (86) R kx h X, R ky h Y (87) Assuming a key exists for V CX. For the first case, consider the following: H(V CX ) H(V X ), H(V CX ) H(V Y ) and H(V CX ) H(V CX ). + log M 2 + log M kcy + log M 3 = 3 log M kcy + log M + log M 2 + log M 3 3h Y ɛ 0 (78) 3h X ɛ 0 h X (79) where (78) results from (59) and the result is from the consideration at the beginning of this proof. K log M ky = K [log M k3 + log M k4 + log M kcy ] = log M kcy + log M 3 + log M kcy + log M 4 + log M kcy = 3 log M kcy + log M 3 + log M 4 3h Y ɛ 0 (80) h Y (8) where (209) results from (59). The security levels thus result: K H(XK W X, W Y ) = K H(X W X W k, (82) W X2 W k2, W CX W k3, W Y W k4, W Y 2 W k5, W CY ) = H(X K ) (83) h X ɛ 0 (84) M CX = 2 Kh X (88) The codewords W X and W Y and their keys W kx and W ky are now defined: W X = (W X W kcx, W X2 W kcx, W CX ) (89) W Y = (W Y W kcx, W Y 2 W kcx, W CY W kcx )(90) W kx = (W kcx ) (9) where W α I Mα = {0,,..., M α }. The wiretapper will not know W X, W X2 and W CX from W X and W Y, W Y 2 and W CY from W Y as these are protected by the key W kcx and W kcx is protected by a random number key. K log M X + K log M Y = K (log M X + log M X2 + log M CX ) + K (log M Y + log M Y 2 + log M CY ) H(X Y ) + H(Y X) + I(X; Y ) + ɛ 0 = H(X, Y ) R X + R Y (92)

4 K log M kx = K log M CX h X ɛ 0 (93) where (93) comes from (88). R kx (94) W Y 2 and W CY from W Y as these are protected by the key (W kcy. In this case, R X, R Y, R kx and R ky satisfy that K log M X + K log M Y = K (log M X + log M X2 + log M CX ) K log M ky = K log M CX = h X (95) h Y (96) R ky (97) where (96) comes from (88) and (96) comes form the consideration stated at the beginning of this proof. The security levels thus result: K H(XK W X, W Y ) = K H(X W X W kcx, W X2 W kcx, W CX, W CY W kcx, W Y W kcx, W Y 2 W kcx ) = H(X K ) (98) h X ɛ 0 (99) where (70) holds because W X, W X2, W CX, W Y, W Y 2 are covered by key W CY and W CY is covered by an existing random number key. Equations (0) - (6) imply that W X, W X2, W Y and W Y 2 have almost no redundancy and they are mutually independent. Similarly, K H(Y K W X, W Y ) h Y ɛ 0 (200) Therefore (R X, R Y, R kx, R ky, h XY, h XY ) is admissible from (09) - (6). Next the case where: H(V CX ) < H(V X ), H(V CX ) < H(V Y ) and H(V CX ) < H(V CX ) is considered. Here, there are shorter length keys used in addition to the key provided by W CX in order to make the key lengths required by the individual portions. For example the key W k comprises W kcx and a short key W, which together are the length of W X. The codewords W X and W Y and their keys W kx and are now defined: W ky W X = (W X W k, W X2 W k2, W CX ) (20) W Y = (W CY W k3, W Y W k4, W Y 2 W k5 ) (202) W kx = (W k, W k2, W k3 ) (203) W ky = (W k4, W k5 ) (204) where W α I Mα = {0,,..., M α }. The wiretapper will not know W X, W X2 and W CX from W X and W Y, + K (log M Y + log M Y 2 + log M CY ) H(X Y ) + H(Y X) + I(X; Y ) + ɛ 0 = H(X, Y ) R X + R Y (205) K log M kx = K [log M k + log M k2 + log M k3 ] = log M kcx + log M + log M kcx + log M 2 + log M kcx + log M 3 = 3 log M kcx + log M + log M 2 + log M 3 3h X ɛ 0 (206) h X (207) where (206) results from (88). K log M ky = K [log M k3 + log M k4 + log M kcy ] = log M kcx + log M 3 + log M kcx + log M 4 + log M kcx = 3 log M kcx + log M 3 + log M 4 3h X ɛ 0 (208) 3h Y ɛ 0 (209) h Y (20) where (208) results from (88) and (209) results from the consideration at the beginning of this proof. The security levels thus result: K H(XK W X, W Y ) = K H(X W X W k, W X2 W k2, W CX, W CY W k3, W Y W k4, W Y 2 W k5, W CY ) = H(X K ) (2) h X ɛ 0 (22) where (2) holds because W X, W X2, W CY, W Y, W Y 2 are covered by key W CX and some shorter length key and W CX is covered by an existing random number key. Similarly, K H(Y K W X, W Y ) h Y ɛ 0 (23) Therefore (R X, R Y, R kx, R ky, h XY, h XY ) is admissible from (205) - (23).

5 B. Converse parts From Slepian-Wolf s theorem we know that the channel rate must satisfy R X H(X Y ), R Y H(Y X) and R X +R Y H(X, Y ) to achieve a low error probability when decoding. Hence, the key rates are considered in this subsection. Converse part of Theorem 2: R kx K logm kx ɛ K H(W kx) ɛ K H(W kx W ) ɛ = K [H(W kx) I(W kx ; W )] ɛ = K [H((W kx X,Y,W ) + I(W kx ; W ) + I(W kx ; X Y, W ) + I(X, Y, W kx W ) + I(Y, W kx X, W ) I(W kx ; W )] ɛ = K [H(X, Y W ) H(X, Y W, W kx )] ɛ h XY K H(X, Y W, W kx ) ɛ (24) = h XY H(V CY ) ɛ = h XY ɛ (25) where (24) results from equation (77). Here, we consider the extremes of H(V CY ) in order to determine the limit for R kx. When this quantity is minimum then we are able to achieve the maximum bound of h XY. R ky K logm ky ɛ K H(W ky ) ɛ K H(W ky W 2 ) ɛ = K [H(W ky ) I(W ky ; W 2 )] ɛ = K [H((W ky X,Y,W 2 ) + I(W ky ; W 2 ) + I(W ky ; X Y, W 2 ) + I(X, Y, W ky W 2 ) + I(Y, W ky X, W 2 ) I(W ky ; W 2 )] ɛ = K [H(X, Y W 2) H(X, Y W 2, W ky )] ɛ h XY K H(X, Y W 2, W ky ) ɛ (26) = h XY H(V CX ) ɛ = h XY ɛ (27) where (26) results from equation (78). Here, we consider the extremes of H(V CX ) in order to determine the limit for R ky. When this quantity is minimum then we are able to achieve the maximum bound of h XY. Converse part of Theorem 3: R kx K logm kx ɛ K H(W kx) ɛ K H(W kx W ) ɛ = K [H(W kx) I(W kx ; W )] ɛ = K [H((W kx X,W ) + I(W kx ; W ) + I(X, W kx W ) I(W kx ; W )] ɛ K I(X, W kx W ) ɛ = K [H(X W ) H(X W, W kx )] ɛ h X H(V CY ) ɛ (28) = h X ɛ (29) where (28) results from (75). Here, we consider the extremes of H(V CY ) in order to determine the limit for R kx. When this quantity is minimum then we are able to achieve the maximum bound of h X. R ky K logm ky ɛ K H(W ky ) ɛ K H(W ky W 2 ) ɛ = K [H(W ky ) I(W ky ; W 2 )] ɛ = K [H((W ky Y,W 2 ) + I(W ky ; W 2 ) + I(X, W ky W 2 ) I(W ky ; W 2 )] ɛ K I(Y, W ky W 2 ) ɛ = K [H(Y W 2) H(Y W 2, W ky )] ɛ h Y H(V CX ) ɛ (220) = h Y ɛ (22) where (28) results from (76). Here, we consider the extremes of H(V CX ) in order to determine the limit for R ky. When this quantity is minimum then we are able to achieve the maximum bound of h Y. Since theorems 4-5 also have key rates of h X and h Y for X and Y respectively we can use the same methods to prove the converse. VI. SCHEME FOR MULTIPLE SOURCES The two correlated source model presented in Section II is generalised even further, and now concentrates on multiple correlated sources transmitting syndromes across multiple wiretapped links. This new approach represents a network scenario where there are many sources and one receiver. We consider the information leakage for this model for Slepian- Wolf coding and thereafter consider the Shannon s cipher system representation.

6 A. Information leakage using Slepian-Wolf coding Here, Figure 5 gives a pictorial view of the new extended model for multiple correlated sources. Figure 5. S S 2 S n T S T S2 T Sn Receiver Extended generalised model Consider a situation where there are many sources, which are part of the S set: S = {S, S 2,..., S n } where i represents the ith source (i =,..., n) and there are n sources in total. Each source may have some correlation between some other source and all sources are part of a binary alphabet. There is one receiver that is responsible for performing decoding. The syndrome for a source S i is represented by T Si, which is part of the same alphabet as the sources. The entropy of a source is given by a combination of a specific conditional entropy and mutual information. In order to present the entropy we first define the following sets: - The set, S that contains all sources: S = {S, S 2,..., S n }. - The set, S t that contains t unique elements from S and S t S, S i S t, S t S c t = S and S t = t Here, H(S i ) is obtained as follows: H(S i ) = H(S i S \Si ) + n ( ) t t=2 all possible S t I(S t S c t)(222) Here, n is the number of sources, H(S i S \Si ) denotes the conditional entropy of the source S i given S i subtracted from the set S and I(S t S c t) denotes the mutual information between all sources in the subset S t given the complement of S t. In the same way as for two sources, the generalised probabilities and entropies can be developed. It is then possible to decode the source message for source S i by receiving all components related to S i. This gives rise to the following inequality for H(S i ) in terms of the sources: n H(S i S \Si ) + ( ) t I(S t S c t) t=2 all possible S t H(S i ) + δ (223) In this type of model information from multiple links need to be gathered in order to determine the transmitted information for one source. Here, the common information between sources is represented by the I(S t S c t) term. The portions of common information sent by each source can be determined upfront and is an arbitrary allocation in our case. For example in a three source model where X, Y and Z are the correlated sources, the common information shared with X and the other sources is represented as: I(X; Y Z) and I(X; Z Y ). Each common information portion is divided such that the sources having access to it are able to produce a portion of it themselves. The common information I(X; Y Z) is divided into V CX and V CY where the former is the common information between X and Y, produced by X and the latter is the common information between X and Y, produced by Y. Similarly, I(X; Z Y ) consists of two common information portions, V CX2 and V CZ produced by X and Z respectively. As with the previous model for two correlated sources, since wiretapping is possible there is a need to develop the information leakage for the model. The information leakages for this multiple source model is indicated in (224) and (225). Remark : The leaked information for a source S i given the transmitted codewords T Si, is given by: L Si T Si = I(S i ; T Si ) (224) Since we use the notion that the information leakage is the conditional entropy of the source given the transmitted information subtracted from the source s uncertainty (i.e H(S i ) H(S i T Si )), the proof for (224) is trivial. Here, we note that the common information is the minimum amount of information leaked. Each source is responsible for transmitting its own private information and there is a possibility that this private information may also be leaked. The maximum leakage for this case is thus the uncertainty of the source itself, H(S i ). We also consider the information leakage for a source S i when another source S j( j i) has transmitted information. This gives rise to Remark 2. Remark 2: The leaked information for a source S i given the transmitted codewords T Sj, where i j is: L Si T Sj = H(S i ) H(S i T Sj ) = H(S i ) [H(S i ) I(S i ; T Sj )] = I(S i ; T Sj ) (225) The information leakage for a source is determined based on the information transmitted from any other channel using the common information between them. The private information is not considered as it is transmitted by each source itself and can therefore not be obtained from an alternate channel. Remark 2 therefore gives an indication of the maximum amount of information leaked for source S i, with knowledge of the syndrome T Sj. These remarks show that the common information can be used to quantify the leaked information. The common information provides information for more than one source and is therefore susceptible to leaking information about more than one source should it be compromised. This subsection gives an indication of the information leakage for the new generalised