Secure Multi-Party Multi-Key Adversarial Cryptography Using GAN

The realm of cryptography, rooted in neural networks, has experienced advancements following the inception of adversarial cryptography. This approach employs Generative Adversarial Networks (GANs) to construct neural networks capable of autonomously acquiring encryption skills. Our primary focus lies in ensuring the confidentiality of information within a multiagent system, with properties articulated in terms of an adversary. Our approach consists of a neural network named `Alice’ to create a ciphertext from plaintext and multiple keys, for data security. Simultaneously, we introduce several neural networks, collectively called `Bob,’ designed to identify and decrypt the ciphertext back to plaintext. Our system is designed to withstand various threats, including intruders possessing knowledge about the encryption process, and potentially having access to leaked ciphertext or a subset of keys. Our approach addresses the challenges posed by attackers engaging in activities like Chosen-Plaintext and Chosen-Ciphertext Attacks. Our system trains `Alice’ and `Bob’ collaboratively to counteract the impact of both intruders and attackers. Over the training process, Alice and Bob reach a state of equilibrium, enabling them to decrypt ciphers in the presence of intruders. Our implementation reveals that nearly 800 training epochs are necessary for model synchronization. Intruders are left with only random guesses about the plaintext, leading to increased errors. Even CPA and CCA attackers fail to strongly predict a probability of either 0 or 1, yielding probabilities close to 0.5. In contrast, Bob decrypts with minimal errors. Moreover, the study illustrates how increasing the number of keys results in faster learning for Bob.