22 May 2002
Analysis of Neural Cryptography
Alexander Klimov, Anton Mityaguine, and Adi Shamir
Computer Science department, The Weizmann Institute, Rehovot 76100, Israel.
{ask,mityagin,shamir}@wisdom.weizmann.ac.il
Abstract. In this paper we analyse the security of a new key exchange protocol proposed in [3], which is based on mutually learning neural networks. This is a new potential source for public key cryptographic schemes which are not based on number theoretic functions, and have small time and memory complexities. In the first part of the paper we analyse the scheme, explain why the two parties converge to a common key, and why an attacker using a similar neural network is unlikely to converge to the same key. However, in the second part of the paper we show that this key exchange protocol can be broken in three different ways, and thus it is completely insecure.
____________________
3. Ido Kanter, Wolfgang Kinzel, Eran Kanter, "Secure exchange of information by synchronization of neural networks'', Europhys., Lett. 57, 141, 2002.
http://cryptome.org/neuralsub.ps (11 pages. 366KB)
This is cryptanalysis of the neural network cryptosystem reported in Bruce Schneier's Crypto-Gram in April 2002:
This is a novel idea. Two neural nets begin with secret random weights and then train on each other's output. Turns out they synchronize much sooner than can an observer who can only see the output but not affect it.