Bereiche | Tage | Auswahl | Suche | Downloads | Hilfe

DY: Dynamik und Statistische Physik

DY 46: Poster

DY 46.6: Poster

Donnerstag, 27. März 2003, 15:30–18:00, P1

Repulsive feedback mechanisms in neural cryptography — •Andreas Ruttor and Wolfgang Kinzel — Theoretische Physik, Universität Würzburg, Am Hubland, 97074 Würzburg

Recently, Kanter et al have shown how to use synchronization of neural networks for cryptography (EPL 57, 141-147, 2002). Two parity machines which learn from each other generate a secret key over a public channel. No prior secret information has to be exchanged. Synchronization is a competition between repulsive and attractive learning steps. An opponent who is listening to the exchange of bits has more repulsive steps. Hence she can synchronize to the secret key with an very low probability, only.

We investigate whether additional repulsive mechanisms improve the security of neural cryptography. A feedback mechanism is constructed by which each network generates a pseudorandom sequence of bits. We show that these bits pass several test of random numbers, hence the neural networks can directly be used for encryption (one-time-pad). When another neural networks is trained on such a sequence, it can learn the weights only if its initial overlap to the generator is very large. Hence this feedback mechanism is strongly repulsive for two interacting networks. We show how to decrease to probability of a successful attack by integrating the feedback mechanism. For publications and preprints see:
http://theorie.physik.uni-wuerzburg/TP3/publi.html