MULTILAYER PERCEPTRON GUIDED KEY GENERATION THROUGH MUTATION WITH RECURSIVE REPLACEMENT IN WIRELESS COMMUNICATION (MLPKG)

Transcription

1 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 MULTILAYER PERCEPTRON GUIDED KEY GENERATION THROUGH MUTATION WITH RECURSIVE REPLACEMENT IN WIRELESS COMMUNICATION (MLPKG) Arndam Sarkar 1 and J. K. Mandal 2 1 Department of Computer Scence & Engneerng, Unversty of Kalyan, W.B, Inda arndam.vb@gmal.com 2 Department of Computer Scence & Engneerng, Unversty of Kalyan, W.B, Inda jkm.cse@gmal.com ABSTRACT In ths paper, a multlayer perceptron guded key generaton for encrypton/decrypton (MLPKG) has been proposed through recursve replacement usng mutated character code generaton for wreless communcaton of data/nformaton. Multlayer perceptron transmttng systems at both ends accept an dentcal nput vector, generate an output bt and the network are traned based on the output bt whch s used to form a protected varable length secret-key. For each sesson, dfferent hdden layer of multlayer neural network s selected randomly and weghts or hdden unts of ths selected hdden layer help to form a secret sesson key. The plan text s encrypted usng mutated character code table. Intermedate cpher text s yet agan encrypted through recursve replacement technque to from next ntermedate encrypted text whch s agan encrypted to form the fnal cpher text through channg, cascaded xorng of multlayer perceptron generated sesson key. If sze of the fnal block of ntermedate cpher text s less than the sze of the key then ths block s kept unaltered. Recever wll use dentcal multlayer perceptron generated sesson key for performng decpherng process for gettng the recursve replacement encrypted cpher text and then mutated character code table s used for decodng. Parametrc tests have been done and results are compared n terms of Ch-Square test, response tme n transmsson wth some exstng classcal technques, whch shows comparable results for the proposed technque. KEYWORDS Multlayer Perceptron, Sesson Key, Encrypton, Mutated Character Code, Wreless Communcaton. 1. INTRODUCTION In recent tmes wde ranges of technques are developed to protect data and nformaton from eavesdroppers [1, 2, 3, 4, 5, 6, 7, 8, 9]. These algorthms have ther vrtue and shortcomngs. For Example n DES, AES algorthms [1] the cpher block length s nonflexble. In NSKTE [4], NWSKE [5], AGKNE [6], ANNRPMS [7] and ANNRBLC [8] technque uses two neural networks one for sender and another for recever havng one hdden layer for producng synchronzed weght vector for key generaton. Now attacker can get an dea about sender and recever s neural machnes because for each sesson archtecture of neural machne s statc. In NNSKECC algorthm [9] any ntermedate blocks throughout ts cycle taken as the encrypted DOI : /jans

2 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 block and ths number of teratons acts as secret key. Here f n number of teratons are needed for cycle formaton and f ntermedate block s chosen as an encrypted block after n/2 th teraton then exactly same number of teratons.e. n/2 are needed for decode the block whch makes easer the attackers lfe. To solve these types of problems n ths paper we have proposed a multlayer perceptron guded encrypton technque n wreless communcaton. The organzaton of ths paper s as follows. Secton 2 of the paper deals wth the problem doman and methodology. Proposed Multlayer Perceptron based key generaton has been dscussed n secton 3. Character code table generaton technque s gven n secton 4. Recursve replacement encrypton and example of encrypton has been presented n secton 5 and 6 respectvely. Secton 7 and 8 deals wth recursve replacement decrypton and example of decrypton method. Complexty analyss of the technque s gven n secton 9. Expermental results are descrbed n secton 10. Analyss of the results presented n secton 11. Analyss regardng varous aspects of the technque has been presented n secton 12. Conclusons and future scope are drawn n secton 13 and that of references at end. 2. PROBLEM DOMAIN AND METHODOLOGY In securty based communcaton the man problem s dstrbuton of key between sender and recever. Because at the tme of exchange of key over publc channel ntruders can ntercept the key by resdng n between them. Ths partcular problem has been addressed and a technque has been proposed technque addressed ths problem. These are presented n secton 2.1 and 2.2 respectvely Man-In-The-Mddle Attack Intruders nterceptng n the mddle of sender and recever and try to capture all the nformaton transmttng from both partes. Dffe-Hellman key exchange technque [1] suffers from ths type of problems. Intruders can act as sender and recever smultaneously and try to steal secret sesson key at the tme of exchangng key va publc channel Methodology n MLPKG Ths well known problem of mddle man attack has been addressed n MLPKG where secret sesson key s not exchanged over publc nsecure channel. At end of neural weght synchronzaton strategy of both partes generates dentcal weght vectors and actvated hdden layer outputs for both the partes become dentcal. Ths dentcal output of hdden layer for both partes can be use as one tme secret sesson key for secured data exchange. 3. MULTILAYER PERCEPTRON BASED KEY GENERATION SYSTEM A multlayer perceptron synaptc smulated weght based undsclosed key generaton s carred out between recpent and sender. Fgure1 shows multlayer perceptron based synaptc smulaton system. Sender and recevers multlayer perceptron select same sngle hdden layer among multple hdden layers for a partcular sesson. For that sesson all other hdden layers goes n deactvated mode means hdden (processng) unts of other layers do nothng wth the ncomng nput. Ether synchronzed dentcal weght vector of sender and recevers nput layer, actvated hdden layer and output layer becomes sesson key or sesson key can be form usng dentcal output of hdden unts of actvated hdden layer. The key generaton technque and analyss of the 12

3 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 technque usng random number of nodes (neurons) and the correspondng algorthm s dscussed n the subsectons 3.1 to 3.5 n detals. Fgure 1. A Multlayer Perceptron wth 3 Hdden Layers Sender and recever multlayer perceptron n each sesson acts as a sngle layer network wth dynamcally chosen one actvated hdden layer and K no. of hdden neurons, N no. of nput neurons havng bnary nput vector, x { 1, + 1}, dscrete weghts, are generated from nput to j output, are les between -L and +L, w j { L, L+ 1,..., + L}.Where = 1,,K denotes the th hdden unt of the perceptron and j = 1,,N the elements of the vector and one output neuron. Output of the hdden unts s calculated by the weghted sum over the current nput values. So, the state of the each hdden neurons s expressed usng (eq.1) 1 1 = (1) N h wx = w, jx, j N N j= 1 Output of the th hdden unt s defned as = sgn( h ) (2) But n case of h = 0 then sum over ts nputs s postve, or else t s nactve, product of the hdden unts expressed n (eq. 2) = -1 to produce a bnary output. Hence a, = +1, f the weghted = -1. The total output of a perceptron s the = K = 1 (3) 3.1 Multlayer Perceptron Smulaton Algorthm Input: - Random weghts, nput vectors for both multlayer perceptrons. Output: - Secret key through synchronzaton of nput and output neurons as vectors. Method:- Step 1. Intalzaton of random weght values of synaptc lnks between nput layer and randomly selected actvated hdden layer. 13

4 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 Where, w j { L L + 1,..., + L}, (4) Step 2. Step 3. Step 4. Step 5. Repeat step 3 to 6 untl the full synchronzaton s acheved, usng Hebban-learnng rules. w j A ( w ( ) ( B + x Θ Θ )) + = g (5),, j, j Generate random nput vector X. Inputs are generated by a thrd party or one of the communcatng partes. Compute the values of the actvated hdden neurons of actvated hdden layer usng (eq. 6) 1 1 = (6) N h wx = w, jx, j N N j= 1 Compute the value of the output neuron usng K = (7) = 1 Compare the output values of both multlayer perceptron by exchangng the system outputs. f Output (A) Output (B), Go to step 3 else f Output (A) = Output (B) then one of the sutable learnng rule s appled only the hdden unts are traned whch have an output bt dentcal to the common output. Update the weghts only f the fnal output values of the perceptron are equvalent. When synchronzaton s fnally acheved, the synaptc weghts are dentcal for both the system. 3.2 Multlayer Perceptron Learnng rule At the begnnng of the synchronzaton process multlayer perceptron of A and B start wth A / B uncorrelated weght vectors w. For each tme step K, publc nput vectors are generated randomly and the correspondng output bts A/B are calculated. Afterwards A and B communcate ther output bts to each other. If they dsagree, A B, the weghts are not changed. Otherwse learnng rules sutable for synchronzaton s appled. In the case of the Hebban learnng rule [10] both neural networks learn from each other. + A ( w ( ) ( )) B + x Θ Θ w j = g, (8), j, j The learnng rules used for synchronzng multlayer perceptron share a common structure. That s why they can be descrbed by a sngle (eq. 4) w j A B ( + f(, ) x ) +, = gw, j,, j (9) 14

5 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 A B wth a functon f (, ),, whch can take the values -1, 0, or +1. In the case of bdrectonal nteracton t s gven by Hebban learnng A B A A B f (,, ) = Θ( ) Θ( ) ant-hebban learnng Random walk learnng 1 (10) A A B A B The common part Θ( ) Θ( ) of f (, ), controls, when the weght vector of a hdden unt s adjusted. Because t s responsble for the occurrence of attractve and repulsve steps [6]. 3.3 Weght Dstrbuton of Multlayer Perceptron In case of the Hebban rule (eq. 8), A's and B's multlayer perceptron learn ther own output. Therefore the drecton n whch the weght w, j moves s determned by the product x, j. As the output s a functon of all nput values, the probabltes to observe x, j = +1 or x, j correspondng weght w, [11, 13, 14, 15, 16]. j 1 ( ) w, j P x, j = 1 = 1 + erf 2 2 NQ w, j Accordng to ths equaton, x, j = sgn( w, j ) x, j and are correlated random varables. Thus = -1 are not equal, but depend on the value of the (11) occurs more often than the opposte, x, j = sgn( w, j ). Consequently, the Hebban learnng rule (eq. 8) pushes the weghts towards the boundares at -L and +L. In order to quantfy ths effect the statonary probablty dstrbuton of the weghts for t s calculated for the transton probabltes. Ths leads to [11]. P ( w = w ), j = P w 0 m = 1 m erf NQ m 1 erf NQ m Here the normalzaton constant 0 P 0 L w = w = L m = erf erf ( m 1) NQ 2 2 s gven by NQ m m 1 ( m 1 ) m 2 2 In the lmt N the argument of the error functons vanshes, so that the weghts stay unformly dstrbuted. In ths case the ntal length of the weght vectors s not changed by the process of synchronzaton. 1 (12) (13) 15

6 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 Q ( t = 0 ) = L ( L + 1 ) 3 (14) But, for fnte N, the probablty dstrbuton tself depends on the order parameter Q Therefore ts expectaton value s gven by the soluton of the followng equaton: L Q = wpw 2, j w= L ( = w) 3.4 Order Parameters In order to descrbe the correlatons between two multlayer perceptron caused by the synchronzaton process, one can look at the probablty dstrbuton of the weght values n each hdden unt. It s gven by (2L + 1) varables. A B ( = a w b) Pa b = P w, j, j =, (16) whch are defned as the probablty to fnd a weght wth w A j (15), = a n A's multlayer perceptron and w B, j = b n B's multlayer perceptron. In both cases, smulaton and teratve calculaton, the standard order parameters, whch are also used for the analyss of onlne learnng, can be calculated as functons of P, [12]. a b L L A 1 A A 2 Q = w w = a P a, b N a = L b = L 1 L L B B B 2 Q = w w = b P a, b N a = L b = L 1 L L AB A B R = w w = abp a, b N a = L b = L Then the level of synchronzaton s gven by the normalzed overlap between two correspondng hdden unts (17) (18) (19) AB = w A w w A A w B w B w B = R Q AB A Q B (20) 3.5 Hdden Layer as a Secret Sesson Key At end of full weght synchronzaton process, weght vectors between nput layer and actvated hdden layer of both multlayer perceptron systems become dentcal. Actvated hdden layer s output of source multlayer perceptron s used to construct the secret sesson key. Ths sesson key s not get transmtted over publc channel because recever multlayer perceptron has same dentcal actvated hdden layer s output. Compute the values of the each hdden unt by 16

7 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 N = sgn wj x j= 1 j sgn( x) 1 f = 0 f 1 f x < 0, x = 0, x > 0. For example consder 8 hdden unts of actvated hdden layer havng absolute value (1, 0, 0, 1, 0, 1, 0, 1) becomes an 8 bt block. Ths become a secret sesson key for a partcular sesson and cascaded xored wth recursve replacement encrypted text. Now fnal sesson key based encrypted text s transmtted to the recever end. Recever has the dentcal sesson key.e. the output of the hdden unts of actvated hdden layer of recever. Ths sesson key used to get the recursve replacement encrypted text from the fnal cpher text. In the next sesson both the machnes started tunng agan to produce another sesson key. Identcal weght vector derved from synaptc lnk between nput and actvated hdden layer of both multlayer perceptron can also becomes secret sesson key for a partcular sesson after full weght synchronzaton s acheved. 4. CHARACTER CODE TABLE GENERATION For plan text tree fgure 2 shows correspondng tree representaton of probablty of occurrence of each character n the plan text. Character t and r occur once and character e occurs twce. Each character code can be generated by travellng the tree usng preorder traversal. Character values are extracted from the decmal representaton of character code. Left branch s coded as 0 and that of rght branch 1. Table 1 shows the code and value of a partcular character n the plan text. From the orgnal tree mutated tree s derved usng mutaton. Fgure 3, 4 and 5 are the mutated trees. After mutaton new code values as obtaned are tabulated n table 2. Tree havng (n-1) ntermedate nodes can generate 2 n-1 mutated trees. In order to obtan unque value, the code length s added to the character f the value s dentcal n the table. (21) 17

8 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 Character Plan text Table1. Code table Code Value t 10 2 r 11 3 e 0 0 Code Table2. Mutated code table Value Code Value Code Value t r e RECURSIVE REPLACEMENT ENCRYPTION Step 1: Decompose the source stream, say, nto a fnte number of blocks, each preferably of the same sze, say, L. Step 2: Calculate the total number of prmes and nonprmes n the range of 0 to (2 L -1). Accordngly, fnd mnmum how many bts are requred to represent each of these two numbers. Step 3 to be appled for all the blocks. Step 3: For the block under consderaton, calculate the decmal number correspondng to that. Say, t s D. Fnd out f D s prme or nonprme. If D s prme, the code value for that block s 1 and f not so, t s 0.In the seres of prmes or nonprmes (whchever be applcable for D) n the range of 0 to (2 L -1), fnd the poston of D. Represent ths poston n terms of bnary values. Ths s the rank of ths block. After repeatng ths step 3 for all the blocks, followng steps are to be followed. Step 4: Say, there are N number of blocks. In the target stream of bts, put all the N code values one by one startng from the MSB poston. So, n the target stream, the frst N bts are code values for N blocks. Step 5: For puttng all the rank values n the target stream, we are to start from the N th bt from the MSB poston and then to come back bt-by-bt. Immedately after the N th bt, put the rank value of the N th block, followed by the rank value of the (N-1) th block, and so on. In ths way, the rank value of the frst block wll be placed at the last. Step 6: Combnng all the code values as well as the rank values, f the total number of bts n the target stream s not a multple of 8, then to make t so, at most 7 bts may have to be nserted. Inserton of these extra bts s to be started from the (N+1) th poston. So, a maxmum of 7 rght shftng operatons may have to be performed n the (N+1) th poston, where that many 0 s are nserted. 18

9 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 Now, MLP (multlayer perceptron) based secret sesson key s use to xor the recursve replacement encrypted stream. Ths MLP secret sesson key s use to xored wth the same length frst ntermedate cpher text block to produce the frst fnal cpher block (MLP secret sesson key XOR wth same length cpher text). Ths newly generated block agan xored wth the mmedate next block and so on. Ths channg of cascaded xorng mechansm s performed untl all the blocks get exhausted. If the last block sze of ntermedate cpher text s less than the requre xorng block sze (.e. weght vector sze) then ths block s kept untouched. 6. EXAMPLE Consder a stream S= of only 16 bts. Apply these steps to obtan the target stream T correspondng to S usng the recursve replacement technque. Decompose S nto four 4-bt blocks takng bts four by four from the MSB, whch are D 1 =1010, D 2 =1001,D 3 =0101 and D 4 =0010. So, as per step 1, L=4. Obtan the total number of prmes n the range of 0 to 2 4-1=15 s 6 (2, 3, 5, 7, 11, 13) and that of nonprmes s 10 (0, 1, 4, 6, 8, 9, 10, 12, 14, 15). To represent the poston of a prme number the number of bts requred s 3, because snce there are 6 prmes, ther postons range from 0 to 5, Smlarly, to represent the poston of a nonprme number the number of bts requred s 4 because ther postons range from 0 to 9 as there are 10 nonprme numbers. Apply the next step 3 for blocks D 1, D 2, D 3 and D 4. The decmal equvalent of D 1 =1010 s 10, whch s the 6 th nonprme. So, the code value of D 1 s C 1 =0 and the rank s R 1 =0110. The decmal equvalent of D 2 =1001 s 9, whch s the 5 th nonprme. So, the code value of D 2 s C 2 =0 and the rank s R 2 =0101. The decmal equvalent of D 3 =0101 s 5, whch s the 2 nd prme. So, the code value of D 3 s C 3 =1 and the rank s R 3 =010. The decmal equvalent of D 4 =0010 s 2, whch s the 0 th prme. So, the code value of D 4 s C 4 =1 and the rank s R 4 =000. To form the target stream, frst we put all the code values one by one startng from the MSB poston to get 0/0/1/1 and they are followed by the rank values of all the blocks startng from the last,.e., 000/010/0101/0110. Here / works just as the separator. Combnng these code values and rank values we obtan , a stream of length 18. To make the length a multple of 8, a block s to be nserted between the code values and the rank values, so that the stream 0011/000000/ s formed. Therefore correspondng to the 16-bt source stream S= , the 24-bt target stream s T= as follows. 19

10 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 Source Stream S= Block wse Decomposton S=1010/1001/0101/0010 D 1 =1010 D 2 =1001 D 3 =0101 D 4 =0010 Code=0 Code=0 Code=1 Code= Insert between Code values and Rank values Target Stream Fgure 6. Pctoral Representaton of Encryptng S= In ths way, we obtan the target stream as T= Now further MLP generated key and recursvely replacement encrypted text s use to fnally encrypt the block. 7. RECURSIVE REPLACEMENT DECRYPTION Durng decrypton, t s to be noted that the recever wll take MLP secret sesson key. Then cascaded xorng operaton s performed usng MLP secret sesson key wth the cpher text. The technque of performng xorng s same that was n encrypton process. Then MLP secret sesson key s use to decpherng the outcomes of the prevous step. Fnally from the outcomes ntermedate encrypted block (E) s extracted and now key s use to decpher the E to get the source stream. 20

11 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 Followng are the set of steps to be followed for the purpose of recursve replacement decrypton: Step 1: Get the unque block length from the key. Say, t s L. Step 2: r Step 3: Calculate the total number of blocks generated from the source stream of bts. The followng does ths calculaton: Total Number of Blocks (B) = Source Stream Sze / Unque Block Length, / denotng the nteger dvson. So, the frst B number of bts, startng from poston 0 (MSB poston) to poston (B -1) n the encrypted stream denotes the code values of B blocks. Calculate the total number of prmes n the range of 0 to (2 L -1). Say, t s P. Hence calculate how many maxmum bts are requred to express P n bnary form. Say, t s X. Then X = log 2 P + 1, where log 2 P denotes the ntegral part of log 2 P. Step 4: Step 5: Step 6: Calculate the total number of nonprmes n the range of 0 to (2 L -1). Say, t s Q. hence calculates how many maxmum bts are requred to express Q n bnary form. Say, t s Y. Then Y = log 2 Q + 1, where log 2 Q denotes the ntegral part of log 2 Q. It s mentonable here that Q = 2 L P. Consder the MSB. It s the code value of the frst source block. If MSB=1, Consder the last block of X bts, convert the bnary number represented by ths block of bts nto the correspondng decmal, Say, t s M. Mark ths block as beng processed. Fnd the M th prme number n the seres of natural numbers (wth the assumpton that the poston of the frst prme number s 0, not 1). The L-bt bnary number correspondng to the decmal prme number obtaned n 2 s the frst source block. Mark the MSB as beng processed. If MSB=0,Consder the last block of Y bts; convert the bnary number represented by ths block of bts nto the correspondng decmal, Say, t s M. Mark ths block as beng processed. Fnd the M th prme number n the seres of natural numbers (wth the assumpton that the poston of the frst prme number s 0, not 1). The L-bt bnary number correspondng to the decmal nonprme number obtaned n 2 s the frst source block. Mark the MSB as beng processed. Repeat step 7 and step 8 for (B-1) number of tmes for the values of I rangng from 1 to (B-1) as there are (B-1) more blocks left to be consdered. Set I = 1. Step 7: Consder the I th bt from the MSB poston. Let t be denoted by T I. If T I = 1, Consder the frst unprocessed block of P bts n the LSB-to-MSB drecton, convert the bnary number represented by ths block of bts nto the correspondng decmal, Say, t s M. Mark ths block beng processed. Fnd the M th prme number n the seres of natural numbers (wth the assumpton that the posto n of the frst prme number s 0, not 1). The L-bt bnary number correspondng to the decmal prme number obtaned n 2 s the I th source block. If T I = 0, Consder the frst unprocessed block of Q bts n the LSB-to- MSB drecton, convert the bnary number represented by ths block of bts nto the correspondng decmal, Say, t s M. Mark ths block beng processed. Fnd the M th nonprme number n the seres of natural numbers (wth the assumpton that the poston of the frst prme number s 0, not 1). The L-bt bnary number correspondng to the decmal nonprme number obtaned n 2 s the I th source block. Step 8: Let I = I

12 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 Step 9: Concatenate all the blocks obtaned so far n the sequence of ther generaton and ths s the source stream. The length of the source stream s (L * B) and accordngly L T (L * B) number of 0 s n the postons between the code values and the target values n the target stream wll reman beng unmarked, as these 0 s were nserted at the end of the encrypton process; L T beng consdered as the length of the target stream. 8. EXAMPLE We contnue wth the same example, where the target stream we obtaned was T = Now, from step 1, from the key we get the unque block length L = 4. Followng step 2, we obtan the total number of blocks as B = 16 / 4 = 4, as t s assumed to be known to the recever that the source stream before beng encrypted was of length 16 bts. Therefore n the encrypted stream, the frst four bts are the code values of four blocks. Followng step 3, we calculate the total number of prmes n the range of 0 to 15 (.e., 2 4 1), whch s P = 6, and to represent t by a bnary number the maxmum number of bts needed s X = 3. Smlarly, followng step 4, we calculate the total number of nonprmes n the range of 0 to 15 (.e., 2 4 1), whch s P = 10, and to represent t by a bnary number the maxmum number of bts needed s Y = 4. Now, followng step 5, we fnd the MSB as 0, so that we are to consder the block of the last Y = 4 number of bts, whch s 0110, the decmal of whch s M = 6. So, we are to fnd the 6 th nonprme number n the seres of natural numbers. It s 10 (assumng that 0 s the 0 th nonprme, 1 s the 1 st nonprme, and so on), the 4-bt bnary of whch s Hence the frst source block s Usng step 6, we can say that step 7 and step 8 are to be repeated for 3 tmes as there are stll 3 blocks left. Step 7 only does the job of movng from one block to another and, n fact, step 8 works n the same way as step 5. So, proceedng n the same way, we obtan the remanng blocks as 1001, 0101 and Followng step 9, we concatenate all the blocks n the same sequence of ther generaton to obtan the source stream S =

13 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 Fgure 7. Pctoral Representaton of Decryptng T = COMPLEXITY ANALYSIS The complexty of the technque wll be O(L), whch can be computed usng followng three steps. Step 1. To generate a MLP guded key of length N needs O(N) Computatonal steps. The average synchronzaton tme s almost ndependent of the sze N of the networks, at least up to N=1000.Asymptotcally one expects an ncrease lke O (log N). Step 2. Complexty of the encrypton technque s O(L). Step Recursve replacement of bts usng prme nonprme recognton encrypton process takes O(L). Step MLP based encrypton technque takes O(L) amount of tme. Step 3. Complexty of the decrypton technque s O(L). Step In MLP based decrypton technque, complexty to convert fnal cpher text nto recursve replacement cpher text T takes O(L). Step Transformaton of recursve replacement cpher text T nto the correspondng stream of bts S = s 0 s 1 s 2 s 3 s 4 s L-1, whch s the source block takes O(L) as ths step also takes constant amount of tme for mergng s 0 s 1 s 2 s 3 s 4 s L-1. So, overall tme complexty of the entre technque s O(L). 23

14 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July EXPERIMENT RESULTS In ths secton the results of mplementaton of the proposed MLPKG technque has been presented n terms of encrypton decrypton tme, Ch-Square test, source fle sze vs. encrypton tme along wth source fle sze vs. encrypted fle sze. The results are also compared wth exstng RSA [1] technque, exstng ANNRBLC [8] and NNSKECC [9].. Table 3. Encrypton / decrypton tme vs. Fle sze Encrypton Tme (s) Decrypton Tme (s) Source NNSKECC Encrypted MLPKG MLPKG NNSKECC Sze (bytes) [9] Sze (bytes) [9] Table 3 shows encrypton and decrypton tme wth respect to the source and encrypted sze respectvely. It s also observed the alternaton of the sze on encrypton. In fgure 8 stream sze s represented along X axs and encrypton / decrypton tme s represented along Y-axs. Ths graph s not lnear, because of dfferent tme requrement for fndng approprate MLP key. It s observed that the decrypton tme s almost lnear, because there s no MLP key generaton process durng decrypton. Encrypton & decrypton tme Source sze Encrypton Decrypton Fgure 8. Source sze vs. encrypton tme & decrypton tme Table 4 shows Ch-Square value for dfferent source stream sze after applyng dfferent encrypton algorthms. It s seen that the Ch-Square value of MLPKG s better compared to the algorthm ANNRBLC [8] and comparable to the Ch-Square value of the RSA algorthm. Stream Sze (bytes) Ch-Square value (TDES) [1] Table 4. Source sze vs. Ch-Square value Ch-Square Ch-Square value n value (MLPKG) (ANNRBLC) [8] Ch-Square value (RSA) [1]

15 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 Fgure 9 shows graphcal representaton of table 4. Fgure 9. Source sze vs. Ch-Square value Table 5 shows total number of teraton needed and number of data beng transferred for MLP key generaton process wth dfferent numbers of nput(n) and actvated hdden(h) neurons and varyng synaptc depth(l). No. of Input Neurons(N) Table 5. Data Exchanged and No. of Iteratons For Dfferent Parameters Value No. of Actvated Hdden Neurons(K) Synaptc Weght (L) Total No. of Iteratons Followng fgure 10. Shows the snapshot of MLP key smulaton process. Data Exchanged (Kb) Fgure 10. MLP Key Smulaton Snapshot wth N=12, K=10 and L=6 25

16 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July ANALYSIS OF RESULTS From results obtaned t s clear that the technque wll acheve optmal performances. Encrypton tme and decrypton tme vares almost lnearly wth respect to the block sze. For the algorthm presented, Ch-Square value s very hgh compared to some exstng algorthms. A user nput key has to transmt over the publc channel all the way to the recever for performng the decrypton procedure. So there s a lkelhood of attack at the tme of key exchange. To defeat ths nsecure secret key generaton technque a neural network based secret key generaton technque has been devsed. The securty ssue of exstng algorthm can be mproved by usng MLP secret sesson key generaton technque. In ths case, the two partners A and B do not have to share a common secret but use ther ndstngushable weghts or output of actvated hdden layer as a secret key needed for encrypton. The fundamental concepton of MLP based key exchange protocol focuses mostly on two key attrbutes of MLP. Frstly, two nodes coupled over a publc channel wll synchronze even though each ndvdual network exhbts dsorganzed behavour. Secondly, an outsde network, even f dentcal to the two communcatng networks, wll fnd t exceptonally dffcult to synchronze wth those partes, those partes are communcatng over a publc network. An attacker E who knows all the partculars of the algorthm and records through ths channel fnds t thorny to synchronze wth the partes, and hence to calculate the common secret key. Synchronzaton by mutual learnng (A and B) s much qucker than learnng by lstenng (E) [10]. For usual cryptographc systems, we can mprove the safety of the protocol by ncreasng of the key length. In the case of MLP, we mproved t by ncreasng the synaptc depth L of the neural networks. For a brute force attack usng K hdden neurons, K*N nput neurons and boundary of weghts L, gves (2L+1)KN possbltes. For example, the confguraton K = 3, L = 3 and N = 100 gves us 3*10253 key possbltes, makng the attack unfeasble wth today s computer power. E could start from all of the (2L+1)3N ntal weght vectors and calculate the ones whch are consstent wth the nput/output sequence. It has been shown, that all of these ntal states move towards the same fnal weght vector, the key s unque. Ths s not true for smple perceptron the most unbeaten cryptanalyss has two supplementary ngredents frst; a group of attacker s used. Second, E makes extra tranng steps when A and B are quet [10]-[12]. So ncreasng synaptc depth L of the MLP we can make our MLP safe. 12. SECURITY ISSUE The man dfference between the partners and the attacker n MLP s that A and B are able to A nfluence each other by communcatng ther output bts & B whle E can only lsten to these messages. Of course, A and B use ther advantage to select sutable nput vectors for adjustng the weghts whch fnally leads to dfferent synchronzaton tmes for partners and attackers. However, there are more effects, whch show that the two-way communcaton between A and B makes attackng the MLP protocol more dffcult than smple learnng of examples. These confrm that the securty of MLP key generaton s based on the bdrectonal nteracton of the partners. Each partener uses a seperate, but dentcal pseudo random number generator. As these devces are ntalzed wth a secret seed state shared by A and B. They produce exactly the same sequence of nput bts. Whereas attacker does not know ths secret seed state. By ncreasng synaptc depth average synchronze tme wll be ncreased by polynomal tme. But success probablty of attacker wll be drop exponentally Synchonzaton by mutual learnng s much faster than learnng by adoptng to example generated by other network. Undrectonal learnng and bdrectonal synchronzaton. As E can t nfluence A and B at the tme they stop transmt due to synchrnzaton. Only one weght get changed where, = T. So, dffcult to fnd weght for attacker to know the actual weght wthout knowng nternal representaton t has to guess. 26

17 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July FUTURE SCOPE & CONCLUSION Ths paper presented a novel approach for generaton of secret key proposed algorthm usng MLP smulaton. Ths technque enhances the securty features of the key exchange algorthm by ncreasng of the synaptc depth L of the MLP. Here two partners A and B do not have to exchange a common secret key over a publc channel but use ther ndstngushable weghts or outputs of the actvated hdden layer as a secret key needed for encrypton or decrypton. So lkelhood of attack proposed technque s much lesser than the smple key exchange algorthm. Future scope of ths technque s that ths MLP model can be used n wreless communcaton. Some evolutonary algorthm can be ncorporated wth ths MLP model to get well dstrbuted weght vector. ACKNOWLEDGEMENTS The author deep sense of grattude to the DST, Govt. of Inda, for fnancal assstance through INSPIRE Fellowshp leadng for a PhD work under whch ths work has been carred out. REFERENCES [1] Atul Kahate, Cryptography and Network Securty, 2003, Tata McGraw-Hll publshng Company Lmted, Eghth reprnt [2] Sarkar Arndam, Mandal J. K, Artfcal Neural Network Guded Secured Communcaton Technques: A Practcal Approach LAP Lambert Academc Publshng ( ), ISBN: , 2012 [3] Sarkar Arndam, Karforma S, Mandal J. K, Object Orented Modelng of IDEA usng GA based Effcent Key Generaton for E-Governance Securty (OOMIG), Internatonal Journal of Dstrbuted and Parallel Systems (IJDPS) Vol.3, No.2, March 2012, DOI : /jdps , ISSN : [Onlne] ; [Prnt]. Indexed by: EBSCO, DOAJ, NASA, Google Scholar, INSPEC and WorldCat, [4] Mandal J. K., Sarkar Arndam, Neural Sesson Key based Trangularzed Encrypton for Onlne Wreless Communcaton (NSKTE), 2nd Natonal Conference on Computng and Systems, (NaCCS 2012), March 15-16, 2012, Department of Computer Scence, The Unversty of Burdwan, Golapbag North, Burdwan , West Bengal, Inda. ISBN , [5] Mandal J. K., Sarkar Arndam, Neural Weght Sesson Key based Encrypton for Onlne Wreless Communcaton (NWSKE), Research and Hgher Educaton n Computer Scence and Informaton Technology, (RHECSIT- 2012),February 21-22, 2012, Department of Computer Scence, Sammlan Mahavdyalaya, Kolkata, West Bengal, Inda. ISBN ,2012 [6] Mandal J. K., Sarkar Arndam, An Adaptve Genetc Key Based Neural Encrypton For Onlne Wreless Communcaton (AGKNE), Internatonal Conf erence on Recent Trends In Informaton Systems (RETIS 2011) BY IEEE, December 2011, Jadavpur Unversty, Kolkata, Inda. ISBN , 2011 [7] Mandal J. K., Sarkar Arndam, An Adaptve Neural Network Guded Secret Key Based Encrypton Through Recursve Postonal Modulo-2 Substtuton For Onlne Wreless Communcaton (ANNRPMS), Internatonal Conference on Recent Trends In Informaton Technology (ICRTIT 2011) BY IEEE, 3-5 June 2011, Madras Insttute of Technology, Anna Unversty, Chenna, Taml Nadu, Inda /11, 2011 [8] Mandal J. K., Sarkar Arndam, An Adaptve Neural Network Guded Random Block Length Based Cryptosystem (ANNRBLC), 2nd Internatonal Conference on Wreless Communcatons, Vehcular Technology, Informaton Theory And Aerospace & Electronc System Technology (Wreless Vtae 2011) By IEEE Socetes, February 28- March 03, 2011,Chenna, Taml Nadu, Inda. ISBN ,

18 Internatonal Journal on AdHoc Networkng Systems (IJANS) Vol. 2, No. 3, July 2012 [9] Mandal J. K., Sarkar Arndam, Neural Network Guded Secret Key based Encrypton through Cascadng Channg of Recursve Postonal Substtuton of Prme Non-Prme (NNSKECC), Internatonal Confference on Computng and Systems, ICCS 2010, November, 2010,Department of Computer Scence, The Unversty of Burdwan, Golapbag North, Burdwan , West Bengal, Inda.ISBN , 2010 [10] R. Mslovaty, Y. Perchenok, I. Kanter, and W. Knzel. Secure key-exchange protocol wth an absence of njectve functons. Phys. Rev. E, 66:066102,2002. [11] A. Ruttor, W. Knzel, R. Naeh, and I. Kanter. Genetc attack on neural cryptography. Phys. Rev. E, 73(3):036121, [12] A. Engel and C. Van den Broeck. Statstcal Mechancs of Learnng. Cambrdge Unversty Press, Cambrdge, [13] T. Godhavar, N. R. Alanelu and R. Soundararajan Cryptography Usng Neural Network IEEE Indcon 2005 Conference, Chenna, Inda, Dec gg [14] Wolfgang Knzel and ldo Kanter, "Interactng neural networks and cryptography", Advances n Sold State Physcs, Ed. by B. Kramer (Sprnger, Berln. 2002), Vol. 42, p. 383 arxv- cond-mat/ , 2002 [15] Wolfgang Knzel and ldo Kanter, "Neural cryptography" proceedngs of the 9th nternatonal conference on Neural Informaton processng(iconip 02).h [16] Dong Hu "A new servce based computng securty model wth neural cryptography"ieee07/2009.j Arndam Sarkar INSPIRE Fellow (DST, Govt. of Inda), MCA (VISVA BHARATI, Santnketan, Unversty Frst Class Frst Rank Holder), M.Tech (CSE, K.U, Unversty Frst Class Frst Rank Holder). Total number of publcatons 8. Jyotsna Kumar Mandal M. Tech.(Computer Scence, Unversty of Calcutta), Ph.D.(Engg., Jadavpur Unversty) n the feld of Data Compresson and Error Correcton Technques, Professor n Computer Scence and Engneerng, Unversty of Kalyan, Inda. Lfe Member of Computer Socety of Inda snce 1992 and lfe member of cryptology Research Socety of Inda. Dean Faculty of Engneerng, Technology & Management, workng n the feld of Network Securty, Steganography, Remote Sensng & GIS Applcaton, Image Processng. 25 years of teachng and research experences. Eght Scholars awarded Ph.D. one submtted and 8 are pursung. Total number of publcatons