Skip to content

J-UNIWARD

Adaptive JPEG Steganography · Holub & Fridrich, WIFS 2013

Embed a message with a Daubechies-8 wavelet distortion cost and watch, side by side, where J-UNIWARD, F5, and naïve LSB place their changes across the image.

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?

Change exposure
Where each method's changes land, ranked against the J-UNIWARD cost map. 0% = hidden in the most textured coefficients; 100% = the smoothest, most conspicuous ones. Lower is stealthier.
DC / flat coefficients hit
Structurally conspicuous edits. J-UNIWARD and F5 never touch them; naïve LSB does.
DCT histogram
The coefficient distribution. F5's magnitude-shrinkage leaves a visible histogram signature.
Detectability label
A rating from placement — "Resistant" through "Detectable" — for the same payload across all three methods.

Embed a message to compare all three methods on the same image and payload.

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

Where changes landed

Terrain: blue = textured (cheap) → red = smooth (costly). Bright dots = changes.

How STC chooses which coefficients to flip (schematic)

The cost map says where hiding is cheap; STC + Viterbi decide which coefficients actually carry the payload. This 3-step schematic shows the idea the real embedder implements.

Step 1 of 3

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: each method's changes are re-projected into the quantized DCT domain (LSB via a real forward DCT) and ranked against the cost map by where they land.

Placement predicts resistance to modern feature-based detectors better than any single first-order test — but it does not prove undetectability under all attacks.

Limitations

  • The placement analysis is a proxy, not a detector. Real-world steganalysis (e.g., SRM, SRNet, XuNet) uses deep learning on rich feature sets.
  • At high payloads even adaptive embedding runs out of textured coefficients — which is why the recommended rate is ≤ 0.3 bpnzac. Above that, all three methods become exposed.
  • 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 — not "undetectable." This is an educational tool, not suitable for adversarial environments.