Text-in-Image Extractor

This Extractor reads text hidden by the Text-in-Image Hider. It reads bits from the lowest bits of the chosen channels, expects a 4-byte length first, and accepts the result only if that length fits inside the image and is at most 5,000,000 bytes. There is no signature, so the only check is that the length makes sense, and then how the bytes look (text, encrypted or noise).

It does not read data hidden by other tools, which use other layouts. For those, try the LSB Extractor with manual settings or the Steganography Analyzer, which shows the bit planes.

Auto-Detect tries 11 settings: Sequential on all channels at 1, 2 and 3 bits, on red and blue at 1 and 2 bits, on green at 1 bit, and Adaptive on all, red and blue at 1 bit. Each row is labelled text, encrypted, noise or empty. There is no signature, so a row is judged by whether its length fits the image and how its bytes read; a length that fits is not proof of a message, and unrelated data could in principle pass. It does not try Scatter, because it cannot guess the seed, and it does not use a password or Reed-Solomon, so an encrypted message is reported as encrypted and has to be decoded again in Manual mode with the password. In our tests a password-protected message of any length was recognised as encrypted from its layout.

Adaptive images made before the edge-map fix are read through a fallback that reproduces the old behaviour and are marked "older edge map" in the list. That restores messages that were written intact; a larger message that the old version had already corrupted when writing cannot be repaired.

If a password was used, run Manual mode with the same settings and the password. If the password is wrong, the payload is not encrypted or the data is damaged, the tool says so in an amber note and shows the undecrypted bytes; it cannot tell you which of the three it was. Only a successful decryption shows the green "AES-256-GCM ✓".

The password is turned into a key with PBKDF2-HMAC-SHA-256 (100,000 iterations) and the salt stored in the payload, and the AES-256-GCM authentication tag then checks it. A wrong password and a damaged payload both simply report failure, so the tool cannot tell you which it was, and there is no way to recover a forgotten password. The full description, including what this scheme does not protect, is on the Text-in-Image Hider page.

The pre-scan shows, for each channel, the p-value of a classic chi-square test on pairs of pixel values (2k and 2k+1). Replacing lowest bits with random-looking bits makes the two counts in each pair nearly equal, which pushes the p-value towards 1; the photographs we measured gave values near 0 before embedding. In our tests on seven real-photo crops the p-value stayed at 0.000 for a text payload at every fill level, and rose to between 0.86 and 1.000 (Sequential, Scatter) only when random-looking data filled the whole image; for Adaptive it did not reliably rise even then. Results vary with the image, the channel and the embedding settings. A low p-value does not prove an image is clean, and a high p-value does not prove that hidden data is present. Use Auto-Detect and the bit planes.