Wednesday, April 04, 2018

[xfkccubv] Decrypt incorrectly to comprehensible text

Use a very good language model (e.g., LSTM) when encoding and encrypting text so that decrypting with the wrong key will result not in nonsense but normal-seeming text.  One could give a false key under duress.

Bit of a problem that it needs to decrypt to text that it is plausible someone would want to encrypt; there isn't much of a language model for that.

Could also do for pictures: there is a good model, namely porn, for pictures people would want to encrypt.  Then, encrypt text as a picture of text.

Another scenario is a master key to a password manager.  Each encrypted password should have a "language model" associated with it (used or valid password characters, or a more sophisticated password generating scheme like diceware), then an incorrect master password will generate incorrect passwords consistent with its language model or scheme.

No comments :