Transformers—Messages in disguise

Compared to human-designed algorithms, Neural Network (NN)-based cryptography can create a new NN-specific cryptographic scheme that changes every time the NN is (re)trained. NN retraining is advantageous because it requires an adversary to restart its process(es) to learn or break the cryptographic scheme every time the NN is (re)trained. However, NN-based encryption faces challenges, including communication overhead due to encoding for bit errors, quantizing the NN’s continuous-valued output, and enabling One-Time Pad encryption. With this in mind, this work introduces the Random Adversarial Data Obfuscation Model (RANDOM), an Adversarial Neural Cryptography (ANC) scheme. RANDOM is computationally efficient, can generate a new cryptographic mapping in 16 seconds, ensures the encrypted message is unique to the encryption key, does not induce any communication overhead, requires around 100 Kb of memory, and provides up to 2.5 Mb/s of end-to-end encrypted communication. • Random Adversarial Data Obfuscation Model (RANDOM) generates a unique encrypted message anytime the key or to-be-encrypted message changes, achieving One-Time Pad (OTP) cryptography. • Random Adversarial Data Obfuscation Model (RANDOM) is a novel neural network-based cryptography architecture capable of generating a new cryptographic mapping within 16 seconds. • Random Adversarial Data Obfuscation Model (RANDOM) introduces three new neural network layers: the projection layer, the inverse projection layer, and the dot-product layer. • Random Adversarial Data Obfuscation Model (RANDOM) requires only 100 kilobytes of memory, making it suitable for deployment by edge devices. • Random Adversarial Data Obfuscation Model (RANDOM) generates a new encryption scheme every time it is reinitialized, enabling Perfect Forward Security (PFS).