Compromising insecure crypto implementations: A deep-learning based cryptosystem-agnostic testing framework
Leonard Hill, Florian Schimanke, Stefan Schiffner, Robert George Mertens · 2024
The use of insecure implementations of cryptographic systems makes encrypted communications vulnerable to practical attacks. Today, attacking, i.e. testing implementations requires human labour and an understanding of the cryptographic system. Automated systematic testing can reduce the insight needed to discover faulty implementations. The approach presented in this paper employs neural networks as the core of a universal framework for cryptographic attacks on arbitrary black-box encryption schemes. The framework trains a neuronal network to automatically perform decryption of ciphertext without knowing the corresponding decryption key. The network approximates the decryption function by encrypting randomly generated plaintext using an arbitrary encryption function and attempting to learn the relationship between plain-and ciphertext. If the decryption function for a certain key is successfully approximated by the framework, the plaintext of any message encrypted with this key can be restored.