Covert and resilient communication over non-cooperative networks
A method, system and apparatus provides network communication between a sender and a receiver in a communication network. At a sender side of the network communication a ciphertext is encoded into a stegotext in a target language using a randomized language model, the ciphertext based upon an original content in a source language. The stegotext randomly generated by the language model to be decoupled from the original content and to have an assigned context in the target language that corresponds to the network communication in the target language. The stegotext is transmitted to a receiver over multiple communication channels of a communication network using one or more applications on an application layer of the communication network.
1 . A method comprising:
at a sender side of a network communication between a sender and a receiver in communication network:
encoding a ciphertext into a stegotext in a target language using a randomized language model, the ciphertext based upon an original content in a source language, the stegotext randomly generated by the language model decoupled from the original content and having an assigned context in the target language that corresponds to the network communication in the target language;
transmitting the stegotext to a receiver over multiple communication channels of a communication network using one or more applications on an application layer of the communication network; and
at a receiver side of the network communication:
inverting the received stegotext to retrieve the original content by the language model decoding the received stegotext into encrypted ciphertext and decrypting the encrypted ciphertext to retrieve the original content, further including:
extracting one or more bit chunks of the encrypted ciphertext;
tracking a conditional probability distribution of each bit chunk in each recurrent step; and
deriving an encoded bit chunk based on the probability score of the token in the stegotext.
2 . The method of claim 1 , where prior to transmitting the stegotext to the receiver, further comprising:
generating a plurality of messages of the stegotext; and
transmitting the plurality of messages of the stegotext to the receiver by distributing the plurality of messages across a plurality of communication channels using the one or more applications.
3 . The method of claim 1 , where prior to transmitting the stegotext to the receiver, further comprising an automatic machine translation (MT) system translating the stegotext into the target language.
4 . The method of claim 1 , where the network communication is a network conversation between the sender and the receiver and responsive to the receiver receiving the stegotext, further comprising at the receiver side of the network conversation:
generating using the language model reply stegotext in the target language, where the reply stegotext is decoupled from the original content and has a context responsive to the stegotext received from the sender in the target language of the network conversation; and
transmitting to the sender the reply stegotext across one or more communication channels of the communication network using the one or more applications.
5 . The method of claim 1 , where the randomized language model converting the ciphertext into the stegotext in the target language includes the randomized language model generating one or more portions of a multi-modal stegotext.
6 . The method of claim 5 , where the one or more portions of the multi-modal stegotext include one or more of text, audio, image and voice.
7 . The method of claim 5 , the randomized language model converting the original content into multi-modal cover content having text and audio responsive to receiving a ciphertext encrypted from the original content and one or more characteristics for an audio portion of the multi-modal stegotext, the text and the audio of the multi-modal stegotext measured in bits.
8 . The method of claim 5 , where the original content is a voice message and prior to encoding the ciphertext into the stegotext, transcribing the voice message to a text message.
9 . The method of claim 1 , the ciphertext generated by encrypting the original content into a random bit string.
10 . The method of claim 9 , further comprising encrypting the original content into the random bit string using a secret key shared by the sender and the receiver.
11 . The method of claim 10 , where the shared secret key is established by a key exchange protocol based on a public key certificate.
12 . The method of claim 1 , where the source language of the original content and the target language of the stegotext are the same language.
13 . The method of claim 1 , where encoding the ciphertext, via the language model, by controlling how a next token is chosen at recurrent steps further comprising:
choosing the next token to be sampled based on a probability distribution determined by the language model and the ciphertext.
14 . The method of claim 13 , where the language model is a byte-level language model or a subword language model that generates a byte in each recurrent step.
15 . The method of claim 1 , the context of the stegotext generated by the language model being coherent to one or more of operational context and dialog context.
16 . The method of claim 1 , the language model applying chatbots to randomly generate the context of the stegotext, the stegotext having randomized frequency and time characteristics.
17 . The method of claim 16 , the user profiles of the sender and receiver of the network communication and metadata of the network communication are obfuscated by the chatbots.
18 . A method comprising:
at a receiver side of a network communication between a sender and a receiver in a communication network:
inverting received stegotext to retrieve original content, the stegotext including a ciphertext encoded into the stegotext in a target language using a randomized language model and the ciphertext based upon an original content in a source language, the inverting including:
the language model decoding the received stegotext into encrypted ciphertext; and
decrypting the encrypted ciphertext to retrieve the original content, further comprising the language model at the receiver side of the network communication:
extracting one or more bit chunks of the encrypted ciphertext;
tracking a conditional probability distribution of each bit chunk in each recurrent step; and
deriving an encoded bit chunk based on the probability score of the token in the stegotext.
19 . The method of claim 18 , where the network communication is a network conversation between the sender and the receiver and responsive to the receiver receiving the stegotext, further comprising at the receiver side of the network conversation:
generating using the language model reply stegotext in the target language, where the reply stegotext is decoupled from the original content and has a context responsive to the stegotext received from the sender in the target language of the network conversation; and
transmitting to the sender the reply stegotext across one or more communication channels of the communication network using the one or more applications.
20 . The method of claim 18 , where the source language of the original content and the target language of the stegotext are the same language.
21 . The method of claim 18 , the context of the stegotext generated by the language model being coherent to one or more of operational context and dialog context.
22 . The method of claim 18 , the language model applying chatbots to randomly generate the context of the stegotext, the stegotext having randomized frequency and time characteristics.
23 . The method of claim 22 , the user profiles of the sender and receiver of the network communication and metadata of the network communication are obfuscated by the chatbots.