Score:-1

Neural Network based on pseudorandom number

pl flag

Recently, I read this paper NEURAL NETWORK BASED CRYPTOGRAPHY. Under the section 3.1 it said:

The aim is to improve the randomness of the random numbers generated by any algorithm using an NN. In order to improve pseudo-random numbers via a neural network, random numbers are generated by a modified subtract with borrow algorithm in MATLAB. The random numbers generated by the modified subtract with borrow algorithm are tested for randomness by NIST. Then, these random numbers are used as input values, initial weight, bias values and the neuron number of hidden layers. The network’s output values are evaluated without training. The output values of the NN are neural network-based pseudo-random numbers. Therefore, the algorithm can be called a neural-based pseudo-random numbers generator (PRNG). The random numbers generated by the NN-based pseudo-random numbers generator are also tested for randomness by NIST.

I was wondering how the network's output values are evaluated without training? If it use the randomness of the input then the structure of networks changes randomly. I was trying to understand the network but couldn't find any clear explanation (like how to build such network) from that paper.

algorithm

Could anyone provide any paper/repository where I can get/idea to building the process of a similar NN (determine Neural network from PRNG or any chaotic/randomness input)?

fgrieu avatar
ng flag
The abstract's _"…pseudo-random numbers are generated and the results are tested for randomness using National Institute of Standard Technology (NIST) randomness tests"_ is enough to disqualify the paper. It should be obvious to anyone with an understanding of crypto that passing a pre-established randomness test is not an interesting security criteria. Nevertheless, that basic error is extremely common in papers made for the sake of making a paper. I recommend against reading this paper, and questionning the competence or/and motivation of anyone recommending otherwise.
emonhossain avatar
pl flag
Thanks for your response @fgrieu. I wasn't know that. I am trying to explore the cryptosystem and come up with my own idea which will be best if AI is involved. Could you suggest any paper which use neural network and probably explain in details?
Maarten Bodewes avatar
in flag
I'd also stop reading after finding out that even common spelling mistakes have not been corrected. That picture alone has "Channal", "Chipper Text" (ciphertext as separate words), "Characres" and "Reciever". If they don't take things seriously, then why should we? Maybe they should write a neural network for detecting bad papers instead, and then feed their own paper into it.
emonhossain avatar
pl flag
I haven't experience to read paper as I recently started to learning. YES, some errors are eye-catchy of that paper but the concept of that paper give me some interest. And I want to see if such cryptosystem was built (yet) using that idea or similar one, @MaartenBodewes. And YES, I am new to cryptosystem with some intermediate knowledge of Machine learning and AI stuff .
fgrieu avatar
ng flag
A widely held belief is that neural networks are not useful to build cryptosystems or RNGs. And breaking cryptosystems using neural networks does not seem to work well either (though I would not rule it out against practical white box cryptography, or more generally AI against crypto with access to side channel, or weak crypto with access to description). Potentially more fruitful: encryption (e.g. fully homomorphic) to protect confidentiality of trained neural networks. An issue is that's a solution in search of problem, but that's the best idea I have for combining neural nework and crypto.
D.W. avatar
fr flag
Cross-posted: https://cstheory.stackexchange.com/q/50538/5038, https://crypto.stackexchange.com/q/95294/351. Please [do not post the same question on multiple sites](https://meta.stackexchange.com/q/64068).
Score:3
fr flag

The paper's proposed scheme is not useful. I don't recommend spending your time on this paper. If you want to generate pseudorandom numbers in practice, either use a standard pseudorandom number generator, or use a cryptographic-strength pseudorandom generator, or use true randomness. There is no reason to use the paper's scheme. We have plenty of standard, well-vetted, time-honored schemes for generating pseudorandom numbers; there is no need for one based on neural networks.

mangohost

Post an answer

Most people don’t grasp that asking a lot of questions unlocks learning and improves interpersonal bonding. In Alison’s studies, for example, though people could accurately recall how many questions had been asked in their conversations, they didn’t intuit the link between questions and liking. Across four studies, in which participants were engaged in conversations themselves or read transcripts of others’ conversations, people tended not to realize that question asking would influence—or had influenced—the level of amity between the conversationalists.