Cryptography, Steganography, and Watermarking: Hiding the Message, Hiding the Fact, Binding the Owner
TL;DR
- Cryptography hides the content of a message. Anyone watching can still see that a secret is being exchanged.
- Steganography hides the existence of a message inside an innocent-looking cover, such as a photo. It fails the moment someone can prove that something is hidden, even if they never read it.
- Watermarking embeds information that is bound to the content. People may know the mark is there; what matters is that it is hard to remove or forge without damaging the content.
- Steganography and watermarking are both forms of data hiding and use many of the same techniques, but they optimize for opposite properties: stealth versus robustness.
- In practice the three are combined: payloads are encrypted before they are hidden, and signatures and watermarks work together to prove where content came from.
Three ways to protect a message
Imagine Alice wants to send something to Bob, and Wendy is watching the channel. There are three very different things Alice might want:
- Wendy should not be able to read it. That is cryptography. Alice sends ciphertext; Wendy sees gibberish.
- Wendy should not even know a message was sent. That is steganography. Alice sends a holiday photo; the message rides inside it.
- Whatever happens to the content, it should still carry Alice’s mark. That is watermarking. Alice publishes a video; even after it is re-encoded, cropped, or re-recorded, a hidden identifier can still be recovered.
These look similar from a distance because all three involve secret keys and transformations of data. But they defend against different adversaries, and a technique that is excellent at one can be useless at another. Petitcolas, Anderson, and Kuhn’s 1999 survey Information Hiding drew the boundaries of this field carefully, and they remain a useful map today.
A very old idea
Hiding messages is far older than encrypting them well. Herodotus describes Histiaeus shaving the head of a trusted slave, tattooing a message on his scalp, and waiting for the hair to grow back before sending him on his way. He also describes Demaratus warning Sparta by writing on the wooden backing of a wax tablet and covering it with fresh wax.
The word itself comes from Johannes Trithemius, whose Steganographia was written around 1499 and published in 1606: a book on cryptography and steganography disguised as a book on magic.
The modern foundations came in stages:
- 1883. Auguste Kerckhoffs argues that a cipher must remain secure even if the enemy knows the system; only the key should be secret.
- 1949. Claude Shannon’s Communication Theory of Secrecy Systems puts cryptography on an information-theoretic footing.
- 1983. Gustavus Simmons formulates the prisoners’ problem, the classic model of steganography.
- 1996. The first Information Hiding workshop is held in Cambridge, bringing steganography, watermarking, and related topics together as one research community.
Cryptography: hiding the content
Cryptography transforms a message so that only someone with the key can recover it. Modern ciphers are designed so that the ciphertext reveals essentially nothing about the plaintext beyond its length. Kerckhoffs’ principle is the norm: algorithms such as AES are public, and security rests entirely on the key.
What cryptography does not hide is the fact of communication. Ciphertext looks random, and random-looking data is itself a signal. A censor who blocks encrypted traffic, or an investigator who flags anyone sending encrypted files, does not need to break the cipher to act.
Cryptography is also deliberately fragile. With authenticated encryption, changing a single bit of the ciphertext makes decryption fail. That is exactly what you want for integrity, and exactly the opposite of what a watermark needs.
Steganography: hiding the existence
The prisoners’ problem
Simmons’ model goes like this. Alice and Bob are in separate prison cells and want to plan an escape. They may exchange messages, but every message passes through Wendy the warden. If Wendy sees anything suspicious, she cuts the channel. Alice and Bob agreed on a secret key before they were locked up.
Encryption alone does not help, because an encrypted message is itself suspicious. Steganography succeeds only if Wendy cannot distinguish Alice’s messages from innocent ones.
Christian Cachin made this precise in 1998. He modelled the warden’s job as a statistical hypothesis test to distinguish objects drawn from the distribution of normal covers from those drawn from the distribution of stego objects. A stegosystem is ε-secure if the relative entropy (Kullback–Leibler divergence) between the two distributions is at most ε. If the two distributions are identical, Wendy can do no better than guessing.
How data gets hidden
- LSB replacement. The simplest method overwrites the least significant bit of pixel values with message bits. The change is invisible to the eye.
- JPEG domain. Tools such as JSteg hid bits in quantized DCT coefficients. F5 (Westfeld, 2001) added matrix embedding, which carries the same message with fewer changes.
- Adaptive embedding. Modern methods such as HUGO (Pevný, Filler & Bas, 2010) and S-UNIWARD (Holub, Fridrich & Denemark, 2014) assign a cost to changing each element and concentrate changes in noisy, textured regions where they are hardest to model, using coding techniques such as syndrome-trellis codes to minimize total distortion.
How it gets caught
Steganalysis is Wendy’s side of the game, and it has repeatedly broken methods that looked invisible.
- Chi-square attack (Westfeld & Pfitzmann, 1999). LSB replacement tends to equalize the counts of pixel values that differ only in their last bit. A simple statistical test on these pairs reveals the embedding.
- RS analysis (Fridrich, Goljan & Du, 2001). Measures how image regularity changes under small flipping operations and can estimate the length of an LSB message.
- Rich models and deep learning. Large sets of hand-designed noise features (Fridrich & Kodovský, 2012) and later convolutional networks such as SRNet (Boroumand, Chen & Fridrich, 2019) detect even adaptive embedding at modest payloads.
Visual invisibility does not guarantee statistical invisibility. Steganography depends on the distributions of cover and stego objects.
Watermarking: binding information to content
Watermarking also hides data inside content, but the threat model flips. The adversary is no longer trying to detect the mark; often everyone knows it is there. The adversary is trying to remove it, forge it, or make it unreadable while keeping the content usable.
Watermarks therefore need robustness. A robust watermark must survive compression, scaling, cropping, filtering, and increasingly, re-generation by AI models. The classic design by Cox, Kilian, Leighton, and Shamoon (1997) spread the mark across the perceptually significant components of an image, the opposite of LSB embedding: if an attacker damages those components enough to remove the mark, they also damage the image.
Not every watermark is meant to survive:
- Robust watermarks carry ownership or tracking information and are designed to survive processing.
- Fragile watermarks are designed to break at the slightest change, so their absence proves tampering.
- Semi-fragile watermarks tolerate benign operations such as compression but break under malicious edits.
Every watermarking system balances three properties that pull against each other: imperceptibility (it doesn’t degrade the content), robustness (it survives processing), and capacity (how many bits it carries). Improving one usually costs another.
A note on “fingerprinting”
The word fingerprinting means two different things in this field. In content identification, it means computing a signature from the content without changing it. In watermarking, it means embedding a unique mark for each recipient so a leaked copy can be traced back to its source. The second meaning has its own theory: collusion-secure codes, introduced by Boneh and Shaw (1998) and made much shorter by Tardos (2003), ensure that when several recipients combine their copies to confuse the mark, at least one of them can still be identified, while innocent users are accused only with negligible probability.
Side by side
| Cryptography | Steganography | Watermarking | |
|---|---|---|---|
| Protects | The content | The fact of communication | The binding between content and an ID |
| Adversary wants to | Read the message | Detect that a message exists | Remove or forge the mark |
| Success means | Ciphertext reveals nothing | Stego is indistinguishable from cover | Mark survives while content stays usable |
| Is existence secret? | No | Yes, that is the point | Usually not |
| Robust to modification? | No (by design) | Rarely | Yes (except fragile marks) |
| Classic attack | Cryptanalysis, key theft | Steganalysis | Removal, desynchronization, collusion |
| Typical use | Secure communication | Covert communication | Ownership, tracing, provenance |
A map of the field
Plotting the hiding techniques on two axes, stealth and robustness, shows why they behave so differently.
Steganography sits high on stealth and low on robustness: a single recompression can destroy the message, and it is built for a channel where Alice controls the file. Fragile watermarks are intentionally neither stealthy nor robust. Visible logos and most robust watermarks give up stealth for robustness; in AI provenance, the existence of the watermark is often announced on purpose.
The top-right corner, both covert and robust, is the hardest place to be. Forensic watermarks used to trace leaked video live there: they must survive camcording and re-encoding, and ideally pirates should not be able to tell that a mark is present, or where it is.
Better together
Real systems rarely use one of these alone.
- Encrypt, then hide. Steganographic payloads are usually encrypted first. Ciphertext is close to random, which helps the embedded bits behave statistically like noise, and it protects the content if the hiding is discovered.
- Sign, then mark. A digital signature proves integrity and who signed, but it sits outside the content, typically in metadata that is easily stripped. A watermark survives but carries few bits. Provenance standards such as C2PA combine them: the signed manifest carries the rich record, and a watermark acts as a soft binding that can look up a stored copy of the manifest after metadata is lost.
- Protect the payload. Forensic watermark payloads are typically protected with error-correcting codes to survive damage, and with keyed or encrypted embedding to make identifiers harder for attackers to read, copy, or forge.
References
- F. A. P. Petitcolas, R. J. Anderson, M. G. Kuhn, “Information Hiding — A Survey,” Proceedings of the IEEE, 87(7), 1999.
- G. J. Simmons, “The Prisoners’ Problem and the Subliminal Channel,” Advances in Cryptology (CRYPTO ’83), 1984.
- C. Cachin, “An Information-Theoretic Model for Steganography,” Information Hiding (IH ’98), LNCS 1525, 1998.
- I. J. Cox, J. Kilian, F. T. Leighton, T. Shamoon, “Secure Spread Spectrum Watermarking for Multimedia,” IEEE Transactions on Image Processing, 6(12), 1997.
- A. Westfeld, A. Pfitzmann, “Attacks on Steganographic Systems,” Information Hiding (IH ’99), LNCS 1768, 2000.
- J. Fridrich, M. Goljan, R. Du, “Reliable Detection of LSB Steganography in Color and Grayscale Images,” ACM Workshop on Multimedia and Security, 2001.
- A. Westfeld, “F5 — A Steganographic Algorithm,” Information Hiding (IH 2001), LNCS 2137, 2001.
- V. Holub, J. Fridrich, T. Denemark, “Universal Distortion Function for Steganography in an Arbitrary Domain,” EURASIP Journal on Information Security, 2014.
- J. Fridrich, J. Kodovský, “Rich Models for Steganalysis of Digital Images,” IEEE Transactions on Information Forensics and Security, 7(3), 2012.
- M. Boroumand, M. Chen, J. Fridrich, “Deep Residual Network for Steganalysis of Digital Images,” IEEE Transactions on Information Forensics and Security, 14(5), 2019.
- D. Boneh, J. Shaw, “Collusion-Secure Fingerprinting for Digital Data,” IEEE Transactions on Information Theory, 44(5), 1998.
- G. Tardos, “Optimal Probabilistic Fingerprint Codes,” STOC, 2003.
- “Data Hiding with Deep Learning: A Survey Unifying Digital Watermarking and Steganography,” arXiv:2107.09287.