Multilayer Perceptron Guided Key Generation Through Mutation with Recursive Replacement in Wireless Communication (MLPKG)

In this paper, a multilayer perceptron guided key generation for encryption/decryption (MLPKG) has been proposed through recursive replacement using mutated character code generation for wireless communication of data/information. Multilayer perceptron transmitting systems at both ends accept an identical input vector, generate an output bit and the network are trained based on the output bit which is used to form a protected variable length secret-key. For each session, different hidden layer of multilayer neural network is selected randomly and weights or hidden units of this selected hidden layer help to form a secret session key. The plain text is encrypted using mutated character code table. Intermediate cipher text is yet again encrypted through recursive replacement technique to from next intermediate encrypted text which is again encrypted to form the final cipher text through chaining, cascaded xoring of multilayer perceptron generated session key. If size of the final block of intermediate cipher text is less than the size of the key then this block is kept unaltered. Receiver will use identical multilayer perceptron generated session key for performing deciphering process for getting the recursive replacement encrypted cipher text and then mutated character code table is used for decoding. Parametric tests have been done and results are compared in terms of Chi-Square test, response time in transmission with some existing classical techniques, which shows comparable results for the proposed technique.

💡 Research Summary

This paper proposes a novel hybrid encryption scheme named Multilayer Perceptron Guided Key Generation through Mutation with Recursive Replacement (MLPKG) for secure wireless communication. The core innovation lies in using the weight synchronization property of multilayer perceptrons (MLPs) to generate session keys without transmitting the key itself over the public channel, combined with a multi-layer encryption process involving mutated character codes and recursive replacement.

The system involves a sender and a receiver, each possessing an identical MLP structure with multiple hidden layers. For each communication session, a single hidden layer is randomly selected and activated, while others remain inactive. Both MLPs receive the same sequence of random binary input vectors. They compute their output bits based on their current internal weights. These output bits are exchanged over the public channel. Using a Hebbian-based learning rule, the weights of the activated hidden layer in both networks are adjusted only when their output bits agree. Through iterative synchronization, the weight vectors of the sender’s and receiver’s MLPs become identical. The synchronized output of the hidden neurons in the activated layer (or the weight vector itself) is then used to form a secret session key.

The encryption process is a three-stage cascade. First, the plaintext is encrypted using a Mutated Character Code Table. This table is generated by creating a Huffman tree based on character frequency in the plaintext and then applying “mutations” to the tree structure, resulting in a non-standard character-to-code mapping. Second, this intermediate ciphertext undergoes Recursive Replacement Encryption. The data is divided into blocks. For each block, a “code” bit is assigned (1 for prime, 0 for non-prime decimal value), and its “rank” (position within the list of all primes or non-primes) is calculated. The final bitstream is constructed by placing all code bits first, followed by all rank bits in reverse order. Third, this second intermediate ciphertext is encrypted using the MLP-generated session key through a Cascaded XOR operation, chaining the XOR result from one block to the next, to produce the final ciphertext.

Decryption is the exact reverse process. The receiver uses its synchronized MLP to generate the identical session key. It reverses the cascaded XOR to obtain the recursive replacement ciphertext, then applies the inverse of the recursive replacement algorithm, and finally decodes the result using the same mutated character code table to recover the original plaintext.

The paper provides a mathematical analysis of the weight synchronization process, including order parameters and weight distribution. Experimental results evaluate the scheme using the Chi-Square test to demonstrate the statistical randomness of the ciphertext and measure response time for transmission. Comparisons with existing classical techniques show that MLPKG offers comparable security and performance.

Key claimed advantages include: (1) Dynamic session key generation without key exchange over the public channel, mitigating man-in-the-middle attacks. (2) Use of a randomly selected hidden layer per session, making the network structure dynamic and harder to attack. (3) A triple-layer encryption scheme (mutation, recursive replacement, key XOR) enhancing confusion and diffusion. The paper concludes by noting the integrative approach of combining neural networks with traditional cryptographic elements and suggests future work on scalability and resistance against more sophisticated attack models.