The initial work was done in:

D. Pointcheval. Les Reseaux de Neurones et leurs Applications Cryptographiques. Technical report, Technical Report, Laboratoire d’Informatique de l’Ecole Normale Superieure, 1995.This scheme was subsequently proven by Klimov

et al. to be particularly vulnerable to genetic algorithms, geometric

considerations, and probabilistic analysis. (Alexander Klimov, Anton Mityagin, and Adi Shamir. Analysis of Neural Cryptography. In ASIACRYPT, pages 288–298, 2002.)

Other references in this area are:

Kinzel, Wolfgang, and Ido Kanter. "Neural cryptography." Neural Information Processing, 2002. ICONIP'02. Proceedings of the 9th International Conference on. Vol. 3. IEEE, 2002.

Kanter, Ido, and Wolfgang Kinzel. "The theory of neural networks and cryptography." Proceedings of the XXII Solvay Conference on physics on the physics of communication. 2002.

Related to this topic is the area of information hiding in neural networks . This concept has been explored in:

Kaili Zhou, Taifan Quan, and Yaohong Kang. Study on Information

Hiding Algorithm Based on RBF and LSB. International Conference

on Natural Computation, 5:612–614, 2008

I also extended these concepts to create a Neural network Trojan which hides the Payload and is activated according to specific inputs. By providing different stopping criteria, the weight changes, and this makes the attack highly polymorphic.

Hope this helps