Steganography

Anyone who wants to keep a message secret has two options. They can make it unreadable, so that a stranger sees that something is there but cannot decipher it. Or they can hide it so well that no one even suspects a message exists. The second path is called steganography. The word comes from Greek and roughly means “covered writing”. A classic example is lemon juice as invisible ink: the sheet looks blank until it is heated.

Hiding instead of scrambling

The difference from encryption is crucial. An encrypted file is immediately noticeable. Anyone monitoring communications sees a jumble of characters and knows: something is being hidden here. In some countries, that alone is enough to make you suspicious. Steganography solves this problem because it never lets the suspicion arise in the first place.

In practice, both methods are combined. The message is first encrypted and the result is then hidden inside an image. If the hiding place is discovered after all, the finder still faces an unreadable file. This combination is considered standard practice, because each method alone has a weak point.

For IT security, this is an uncomfortable topic. Malware uses steganography to smuggle commands inside seemingly harmless image files. Filters that watch for suspicious file types then fail to catch anything. Conversely, journalists and human rights groups use the technique to stay in contact within surveillance states.

How a message fits inside an image

A digital photo consists of millions of pixels. Each pixel stores its color as a number, usually between 0 and 255. If you change this number by one, the hue shifts minimally. No eye can see the difference between a red value of 174 and 175. It is exactly this lowest digit of each number that can be overwritten.

This method is called the LSB technique, after the English “least significant bit”, i.e. the least important digit. This way, a tiny part of the message can be fitted into each pixel. A photo with a million pixels can thus hold several hundred thousand characters of text. The recipient reads out only the lowest digits and assembles the message from them.

There are other ways too. Data can be stored in unused areas of a file, in audio tracks beyond the audible range, or in the choice of certain words in a text. This can be uncovered through steganalysis: experts compare the statistical distribution of the numbers with that of an unaltered photo. Hidden data often leaves behind a strikingly uniform pattern there.

Watermarks and AI-generated texts

Today, the technique is most commonly encountered as a digital watermark. Film studios and music publishers embed invisible identifiers into their files. If a pirated copy surfaces, the identifier reveals its source. Authorities also mark documents this way to track down leaks.

In AI news, steganography has recently taken on a new role. Because images and texts from chatbots and image generators can hardly be distinguished from real ones anymore, they are supposed to carry a hidden marker. For images, pixels are altered; for texts, word choice is subtly steered in an agreed-upon direction. A detection program recognizes the pattern, a reader does not.

A common misconception is that such markings are tamper-proof. They are not. Anyone who heavily shrinks an image, recompresses it, or has a text rephrased often destroys the watermark. Steganography protects against accidental discovery, not against targeted attacks. This is precisely what is being debated when lawmakers call for mandatory labeling of AI content.