J-UNIWARD Steganography Lab

Adaptive JPEG steganography that reduces statistical detectability by concentrating changes in perceptually complex regions.

Compare J-UNIWARD against naive methods like LSB and F5. See how adaptive placement improves resistance under included statistical tests — all computed locally in your browser.

Start Here

1
Load an image

Use the sample image or upload your own JPEG.

2
Enter a message & key

Type a short secret message and a shared key.

3
Embed with J-UNIWARD

Click Embed and watch the analysis panel light up.

→ See changes concentrated in textured regions

→ Compare detection scores across three methods

Cover Image

📷 Drop a JPEG here or click to browse — JPEG only

Embed / Extract

0 chars · 0 bytes
0.05 (safe) 0.25 (moderate) 0.50 (risky)
Presets:

Upload a stego JPEG, enter the key, and click Extract. If you just embedded above, click Extract without uploading.

Steganalysis Comparison

What will this panel show?

Chi-square test
Detects non-random patterns in DCT coefficient pairs. Lower p-value = more detectable.
Histogram analysis
Visualizes the distribution of DCT coefficients. Embedding distorts the natural shape.
p-value
Probability of seeing this histogram by chance. Below 0.05 = statistically significant tampering.
Detectability label
A qualitative rating based on p-value and change rate — from "Resistant" to "Trivially Detectable."

Embed a message to see live results from all three methods.

Load an image and embed a message to see steganalysis results.

DCT Block Change Map

Red = modified blocks · Gray = unchanged

Methodology

  • Cost function: Daubechies-8 three-level wavelet decomposition measures how much a ±1 DCT change disturbs the image in 9 detail subbands.
  • Embedding: Syndrome-Trellis Codes (h=12, 4096 states) find the minimum-distortion modification via Viterbi search.
  • Analysis: Chi-square test on adjacent DCT histogram pairs (Pairs-of-Values attack) — a first-order statistical test.

This analysis measures first-order histogram distortion only. It does not prove undetectability under all attacks.

Limitations

  • This demo uses simplified, first-order steganalysis. Real-world detectors (e.g., SRNet, XuNet) use deep learning on rich feature sets.
  • The COM marker sideband for salt/rate may be stripped by image pipelines, social media compression, or metadata-stripping tools.
  • J-UNIWARD is more resistant than LSB/F5 under included tests — not "undetectable" or "invisible to analysis."
  • This is an educational tool. It is not suitable for adversarial environments.