|
Mathematical Foundations of Computer Security
Neural network obfuscation for computations over encrypted data
V. L. Eliseevab a Infotecs, Moscow
b National Research University "Moscow Power Engineering Institute"
Abstract:
An approach to neural network cryptographic obfuscation of computations is proposed. Applying the previously obtained results on the property of strict obfuscation of indistinguishability for a neural network approximator, we propose to use neural networks to perform arithmetic and other operations on encrypted data, thus realizing the idea of using homomorphic encryption to perform trusted computations in an untrusted environment. The cryptographic properties of this mechanism are evaluated and compared with traditional approaches to encryption based on the secret key. The advantages and disadvantages of neural networks in relation to the problem of obfuscation and processing of encrypted data are discussed.
Keywords:
artificial neural network, obfuscation, homomorphic encryption, secrecy estimation.
Citation:
V. L. Eliseev, “Neural network obfuscation for computations over encrypted data”, Prikl. Diskr. Mat. Suppl., 2020, no. 13, 85–93
Linking options:
https://www.mathnet.ru/eng/pdma505 https://www.mathnet.ru/eng/pdma/y2020/i13/p85
|
|