Implementing Cryptography in AI Systems

Researchers are exploring how to integrate cryptography into deep neural networks (DNNs) to enable functionalities like decrypting encrypted inputs or verifying authorized access. This is challenging because traditional cryptographic methods are designed for digital computers that process binary data, whereas DNNs operate on continuous real numbers using linear mappings and ReLUs. The difference between discrete and continuous computational models raises questions about the best way to implement standard cryptographic primitives as DNNs. It is also unclear whether DNN-based cryptosystems remain secure when an attacker can input arbitrary real numbers. The goal is to develop a secure way to implement cryptographic primitives in DNNs. This would allow DNNs to perform tasks like decrypting encrypted data or verifying authorized access. The research aims to address the discrepancy between traditional cryptography and DNNs' continuous computational model. The implementation of cryptographic primitives in DNNs must ensure security in the face of potential attacks. The study's outcome could have significant implications for the secure deployment of DNNs in various applications. By exploring this intersection of cryptography and DNNs, researchers can develop new methods for securing DNN-based systems.