Research Explores Secure Cryptography Implementation in Deep Neural Networks

A recent study titled "How to Securely Implement Cryptography in Deep Neural Networks" examines the integration of cryptographic functionalities into deep neural networks (DNNs). DNNs, being analog computers, differ from traditional digital computers, which presents challenges for implementing cryptographic primitives. The study demonstrates that natural implementations of block ciphers as DNNs can be broken in linear time using non-standard inputs. The researchers tested this attack on AES-128 and successfully found randomly chosen keys. They then propose a new method to securely and correctly implement any desired cryptographic functionality as a ReLU-based DNN with low overhead.