Steganography and Watermarking

SteganographySearch on YouTubeBookmarkSaved

Steganography is the practice of concealing information inside another medium (text, image, audio or video) in a way that hides its existence.

• A common example hides a message in the pixels of an image or in the audio of a file, with no noticeable change.
• It is used in cybersecurity to transmit sensitive data covertly or to evade detection.

Examples: a hidden message in an image, an audio file with concealed text, data embedded in a video, and text steganography.

Key principle: statistical undetectability. The stego file must not differ statistically from a normal file.

Steganography versus watermarking (first look)Search on YouTubeBookmarkSaved

Both embed information in digital media, but for different purposes.

• Steganography conceals the existence of information.
• Watermarking embeds identifiable information (a logo, serial number or copyright notice) to show ownership or authenticity.

The full comparison is in the watermarking section.

How it startedSearch on YouTubeBookmarkSaved

Building blocks of steganographySearch on YouTubeBookmarkSaved

1. The choice of cover work
2. The embedding and extracting algorithms
   • Symbol assignment function
   • Embedding modification
   • Selection rule
3. Stego key management

Cover workSearch on YouTubeBookmarkSaved

The cover work is the medium that hides the secret: an image, audio file, video or text. It must look natural after embedding so nobody suspects anything. A complex image hides data better than a plain one.

Embedding and extracting algorithmsSearch on YouTubeBookmarkSaved

These are the steps that hide (embed) the data in the cover and later retrieve (extract) it.

• Symbol assignment function: decides how the secret is broken into small units (symbols or bits) and assigns them to parts of the cover.
• Embedding modification: changes the cover to include the hidden data, with slight, usually imperceptible changes.
• Selection rule: decides which parts of the cover are changed.

Three embedding principlesSearch on YouTubeBookmarkSaved

Data masking is a further approach. It modifies the cover's statistical properties so the hidden message looks like a typical signal (a song or an image), avoiding anomaly-checking tools.

Selection rule in cover modificationSearch on YouTubeBookmarkSaved

Modifications must be as subtle as possible. Where the data goes matters:

• Sequential: in order, pixel by pixel.
• Pseudo-random: at random locations chosen with a key.
• Adaptive: locations chosen by file properties, such as noisy image areas.

Stego key managementSearch on YouTubeBookmarkSaved

A stego key is like a password that controls embedding and extraction. It is needed to access the hidden information.

Tools to explore: OpenStego, SpamMimic (https://www.spammimic.com/index.cgi), Steghide.

LSB embeddingSearch on YouTubeBookmarkSaved

LSB embedding replaces the least significant bit of each selected pixel with one message bit. Each pixel value changes by at most 1, which the eye cannot see.

Cachin's definition of steganographic securitySearch on YouTubeBookmarkSaved

Challenge: hiding is not enough. If a spy or investigator spots the pattern or the hidden message, the whole idea fails.

According to Cachin, the hidden message is truly secure if nobody can detect its presence, even with advanced tools or statistical analysis.

Practical steganographic methodsSearch on YouTubeBookmarkSaved

1. Statistics-preserving steganography
2. Model-based steganography
3. Masking embedding as natural processing

Statistics-preserving steganographySearch on YouTubeBookmarkSaved

Embedding changes statistics, and analysts run tools on patterns, distributions and frequency counts to find tampering. Statistics-preserving methods keep the cover's statistical properties (for example the histogram) the same after embedding.

OutGuessCopy Anchor Link

OutGuess embeds secret messages in JPEG images. Its main advantage is that it preserves the image statistics, so common steganalysis cannot detect it. It works in two phases.

Limitations of OutGuess:

• Over-embedding risk: a large message makes perfect statistical consistency hard, so detection gets easier.
• File size constraints: works best on large, noisy images (landscapes). It can fail on small or uniform images where changes show.

OutGuess versus Steghide (slide table)Copy Anchor Link

Model-based steganographySearch on YouTubeBookmarkSaved

Statistics-preserving methods keep specific metrics (such as histograms). Model-based steganography keeps the conceptual model of the cover intact.

• Split the cover into an invariant part $C_{inv}$ and an embedding part $C_{emb}$.
• Keep the features of $C_{inv}$ (histograms, patterns) unchanged; modify only $C_{emb}$.
• Key principle: embed so that the modified data still fits the same probabilistic model as the original.
• The designer chooses a cover model expressed as conditional probabilities $P(c_{emb} \mid c_{inv})$.
• These probabilities are computed from the invariant part, so sender and receiver hold the same information.

A probabilistic model of an image describes how likely pixel values or features are. It predicts the expected colours, brightness or frequencies from the pixel distribution.

When embedding, keep the same probability structure, for example by changing low-frequency components only slightly.

Masking embedding as natural processingSearch on YouTubeBookmarkSaved

The third method makes the embedding changes look like normal processing of the cover (the deck lists it by name only).

SteganalysisSearch on YouTubeBookmarkSaved

Steganalysis detects hidden messages. The classic scenario has Alice (sender), Bob (receiver) and a Warden in between.

Types of wardensSearch on YouTubeBookmarkSaved

Detection as classificationSearch on YouTubeBookmarkSaved

Every steganalysis model classifies the input as cover or stego. The difficulty is the warden's knowledge.

• Passive wardens know little or nothing about the stego process. They need an algorithm that detects all sorts of steganography: blind steganalysis.
• Malicious wardens know more, including the algorithm. They can use targeted steganalysis.

A classifier makes decisions from parameters. In steganalysis it classifies a file by its characteristics, called features.

• Challenge: the exact distributions of cover and stego works are unknown. So use features: specific parts of the data that give clues.
• Feature selection is crucial. Good features separate cover from stego with fewer errors.
• Classification accuracy: ideally covers and stegos form separate groups in feature space. No overlap means perfect classification. In practice they overlap, which causes errors. Choose features that group covers tightly and keep stegos far away.

A feature is a measurable property of a file that helps separate stego works from cover works. Examples: histogram of pixel intensity, texture patterns, Fourier transform features.

Guidelines for choosing features (trial and error):

• Capture subtle changes.
• Avoid features that change naturally.
• Use feature dimensionality (multiple features).

Targeted steganalysisSearch on YouTubeBookmarkSaved

In targeted steganalysis you know how the message was embedded and which algorithm was used. So you focus on features that the algorithm is most likely to disturb.

Example: LSB embedding. Changing LSBs causes no visible difference, but it adds noise. That noise increases the absolute difference between neighbouring pixels. So the classifier uses the sum of absolute differences between neighbours as its feature.

Blind steganalysisSearch on YouTubeBookmarkSaved

Blind steganalysis must detect the hidden message without knowing the algorithm. This is harder.

Is there one feature that detects every stego method? No. A perfect feature set does not exist. If it did, steganographers would design new algorithms to bypass it.

How to train the classifier:

Universal blind steganalysis: blind classifiers designed to detect steganography in any domain (spatial or JPEG) are universal classifiers. Features: noise patterns, pixel correlation, transform-domain features.

WatermarkingSearch on YouTubeBookmarkSaved

Watermarking: embedding information in a host signal (image, video, audio) so that it is imperceptible but detectable or extractable.

Purpose: protect intellectual property, enforce copyright and prove ownership.

Steganography versus watermarkingSearch on YouTubeBookmarkSaved

Verdict: the goals differ, but many techniques are similar or identical.

Importance of digital watermarkingSearch on YouTubeBookmarkSaved

Concern over copyright in the digital age drives interest in watermarking.

• The internet as a distribution platform: owners face high risk from easy unauthorized distribution.
• High-capacity digital recording devices: near-perfect copies with little or no quality loss make it easy for pirates to spread high-quality illegal copies.

Why not cryptography?Search on YouTubeBookmarkSaved

• First line of defence: encryption makes sure only buyers with the right key can access the content.
• Limitation: cryptography protects content only in transit. Once decrypted, nothing stops illegal distribution.

Need for watermarkingSearch on YouTubeBookmarkSaved

• Complement to cryptography: protection continues after decryption.
• Robustness: watermarks survive decryption, re-encryption, compression, digital-to-analog conversion and format changes.
• Potential solution: the watermark sits in the content and is never removed during normal use, so protection is long-term.

ApplicationsSearch on YouTubeBookmarkSaved

Copy prevention, copyright protection, broadcast monitoring, transaction tracking, owner identification, proof of ownership, content authentication.

Properties of watermarking systemsSearch on YouTubeBookmarkSaved

The weight of each property depends on the application and the watermark's role.

Embedding propertiesSearch on YouTubeBookmarkSaved

• Effectiveness: the watermark serves its purpose (copyright protection, authentication).
• Fidelity: the watermark does not noticeably degrade the content; it is imperceptible.
• Payload: the amount of information the watermark carries. Higher payload can hurt fidelity and robustness.

Detection propertiesSearch on YouTubeBookmarkSaved

• Blind detection: detect without the original, unwatermarked content.
• Informed detection: needs the original content to verify the watermark.
• False positive behaviour: minimize detecting a watermark where none exists.
• Robustness: survive compression, format conversion and digital-to-analog conversion.

Other propertiesSearch on YouTubeBookmarkSaved

• Modifiability and multiple watermarks: the ability to change the encoded information, by altering the watermark or by using several watermarks.
• Costs: embedding cost (time, computation, complexity to embed) and detection cost (resources to detect or extract; matters for real-time or large-scale use).
• Security: resistance to hostile attacks, meaning any process meant to defeat the watermark's purpose. Three kinds: unauthorized removal, unauthorized embedding and unauthorized detection.

Evaluation of watermarking systemsSearch on YouTubeBookmarkSaved

Systems are evaluated to judge fit for an application or to show algorithmic improvement. Key criteria: application suitability and performance metrics (robustness, security, fidelity).

The notion of "best"Search on YouTubeBookmarkSaved

First define what makes one system better than another. "Best" depends on the application.

• Application-specific criteria: in a video watermark for copy control, robustness against small rotations may be critical, because rotations are used as attacks. That makes it the optimal choice for broadcast monitoring.
• Flexibility with new systems: an improvement in a single property (robustness, payload or fidelity) can make a new system valuable where that property matters.

Trade-offs between propertiesSearch on YouTubeBookmarkSaved

Conclusion: the best system depends on the application. Evaluation needs a careful understanding of trade-offs between robustness, payload, fidelity and false positives.

BenchmarkingSearch on YouTubeBookmarkSaved

Benchmarking evaluates and compares watermarking systems: do they meet application needs, or show algorithmic improvement?

• Application-specific testing: each criterion is defined by minimum requirements. Example: the Copy Protection Technical Working Group (CPTWG) tested watermarks for DVD copy control.
• Universal benchmark: a scalar score to compare systems. In an image benchmark, systems must embed a specific payload (for example 80 bits) at a set fidelity.
• Limitations of universal benchmarks: not suitable for all systems (priorities such as robustness and false positive rate differ); application-specific security is not fully evaluated; no single benchmark is universally relevant.
• Conclusion: tailor benchmarking to the application or the improvement being studied.

Models of watermarkingSearch on YouTubeBookmarkSaved

Watermarking as a communication problemSearch on YouTubeBookmarkSaved

There are three communication-based models. Each treats the original content differently.

1. Basic model
2. Watermarking with side information
3. Watermarking as multiplexed communications

Basic modelSearch on YouTubeBookmarkSaved

The original content $c_o$ is treated as noise that interferes with the watermark. The encoder turns message $m$ into an added pattern $w_a$ using the watermark key. The watermarked work is $c_w = c_o + w_a$. Channel noise $n$ gives $c_{wn}$. The detector, with the same key, outputs $m_n$.

The deck's example uses E_BLIND: a reference pattern $w_r$ is added to the original image to give the watermarked image.

Watermarking with side informationSearch on YouTubeBookmarkSaved

In many cases the encoder knows the original content. Side information is that extra knowledge of the original before the watermark is added. With it, the encoder blends the watermark into the content better.

Without side information, the reference pattern is applied uniformly, so it may show in some regions. With side information:

• Dark areas: increase watermark intensity. The eye notices changes in dark areas less.
• Bright areas: reduce intensity so the watermark is not visible.

Watermarking as multiplexed communicationsSearch on YouTubeBookmarkSaved

Two messages share one medium:

• Cover work $c_o$: the content meant for human perception.
• Watermark message $m$: hidden from humans, decoded by machines.

The embedding must keep the watermark hidden from human perception but detectable by the machine.

Robust watermarking approachesSearch on YouTubeBookmarkSaved

Some strategies aim at robustness to all distortions. Others handle specific distortions.

Redundant embeddingSearch on YouTubeBookmarkSaved

The same watermark is embedded in multiple parts or coefficients. Against attacks such as cropping and filtering, it survives in the unaffected areas.

Spread spectrum codingSearch on YouTubeBookmarkSaved

The watermark signal is spread over a wide range of frequencies or components. Damage to some parts leaves it detectable in the rest. The idea comes from communication systems, where a message is spread over many frequencies against loss and interference.

Embedding in perceptually significant coefficientsSearch on YouTubeBookmarkSaved

Embed in the parts most likely to stay unchanged during normal processing: the perceptually significant coefficients that hold information important to human perception (mid-range audio frequencies, important visual textures).

Why these spots?

• Frequency coefficients are not equally stable.
• High-frequency components are easily damaged by compression or filtering.
• Low-frequency components are more stable but more noticeable, so the watermark would be visible.
• Hence target mid-frequency coefficients.

154   5   3   0   0   0   0   0
 10  25   2   1   0   0   0   0
  8   7  50  15   3   0   0   0
  0   4  18  48  22   0   0   0
  0   0   8  30  45  10   0   0
  0   0   0   2  15  55  20   5
  0   0   0   0   5  20  48  10
  0   0   0   0   0   5  10  25

Embedding in coefficients of known robustnessSearch on YouTubeBookmarkSaved

1. Choose a robust domain: one where the expected processing does little harm.
   • Expected spatial shifting (small translations or rotations): embed in the Fourier magnitudes.
   • Expected linear filtering (blurring): embed in the DCT.
2. Identify robust coefficients in that domain, by an analytical approach or empirical testing.
3. Statistical analysis of coefficients: run statistical tests over a large number of works to list coefficients that perform well across content types.

Challenges remain with compression and thresholding.

# embed
steghide embed -cf scene.jpg -ef secret.txt -sf steghide_scene.jpg -p lab123
outguess -k lab123 -d secret.txt scene.jpg outguess_scene.jpg
# extract
steghide extract -sf steghide_scene.jpg -p lab123
outguess -k lab123 -r outguess_scene.jpg secret.txt
# steganalysis
stegdetect -t jopifa cover/scene.jpg stego/*.jpg
cmp -l scene.bmp steghide_scene.bmp | head -3   # byte offsets that changed

Quick self-testSearch on YouTubeBookmarkSaved