Hiding Secrets in Recipe Blogs with Neural Linguistic Steganography

In a digital age where privacy is scarce and data is constantly harvested, I explore a whimsical yet clever method to conceal information: embedding it within the mundane text of recipe blogs. Using a Python CLI tool, I encode messages into seemingly innocuous recipe introductions, leveraging neural linguistic steganography. This technique transforms secret data into natural language, hiding it in plain sight where AI scrapers and nosy agencies might overlook it.

The process involves converting a secret message into a binary fraction, which is then encoded using a model's token probability distribution. By selecting tokens that correspond to specific intervals on a number line, I can embed the message within the text. For example, a message like 'https://www.nokings.org/' can be hidden in a recipe for 'One-Pan Garlic Butter Chicken with Herbed Potatoes'.

Decoding the message requires the same model and prompt, ensuring that the probability distributions align perfectly. However, this method is not without its limitations. It is computationally expensive, as it requires large models to encode and decode small amounts of data. Additionally, issues like BPE tokenization can cause discrepancies if tokens do not re-tokenize consistently.

To mitigate some of these challenges, I introduce solutions like filtering out tokens that wouldn't survive a round-trip through decoding and tokenization. Despite these hurdles, the technique remains a playful way to pass secret messages, akin to passing notes in class.

Security-wise, the prompt acts as a shared key, but it's not foolproof since the generated text is influenced by the prompt itself. While primarily designed for local use, the method could potentially be adapted for remote APIs, provided they offer deterministic outputs.

For those interested in experimenting with this quirky form of steganography, the tool is available on Google Collab and GitHub. So, why not have some fun cooking up secret messages?