Steganalysis toolkit
Defensive research on hidden information
A toolkit for measuring how fragile or resilient hidden-information schemes are (LSB, invisible Unicode, visible watermarks) across text and images.
Project brief
ShippedProblem
Hidden-information schemes are hard to compare without one repeatable way to detect, modify, and verify them.
Outcome
A toolkit for measuring how fragile or resilient hidden-information schemes are (LSB, invisible Unicode, visible watermarks) across text and images.
Evidence
- Graduate research on image steganalysis, feature engineering and model comparison with held-out evaluation, not production deployment.
Key decisions
- Detection zone: PNG, LSB, Unicode, watermark, and optional model-assisted analyzers share one substrate.
- Loop zone: Agentic, strategy-ladder, and multi-role paths all call the same verify contract.
- Cleaning zone: Six backends remove or sanitize detected signals with adaptive escalation on retry.
- Defensive scope: Framing stays defensive; defeating provenance standards like C2PA is explicitly out of scope.
How it works
Hidden-information schemes are hard to compare without one repeatable way to detect, modify, and verify them. I built the toolkit around one reusable loop: detect, modify, verify, retry, with a single shared verification contract so every detector and cleaner speaks the same language. The framing is defensive throughout: the point is understanding robustness, and the scope explicitly rules out defeating provenance standards like C2PA. Six detectors, six cleaners, and an optional CNN tier. This is graduate research on image steganalysis (feature engineering and model comparison with held-out evaluation), not a production deployment. Hidden information breaks is the essay form of the same measurement posture. Measurement as gate is the same posture outside steganography. It shares evaluation rigor with Agentic Eval Plugin.
Architecture
One verify contract underlies both figures: the detect, modify, verify loop that does the actual work, and the toolkit layout arranged around it.
Step 3 · Verify (shared contract) is the gate in this figure: dim until the shared verify contract passes, lit once it does.
Step 05 · Verify (shared contract) is the gate in this figure: dim until the shared verify contract passes, lit once it does.
- Detection zone
- PNG, LSB, Unicode, watermark, and optional model-assisted analyzers share one substrate.
- Loop zone
- Agentic, strategy-ladder, and multi-role paths all call the same verify contract.
- Cleaning zone
- Six backends remove or sanitize detected signals with adaptive escalation on retry.
- Defensive scope
- Framing stays defensive; defeating provenance standards like C2PA is explicitly out of scope.
Skills applied
- ML
- Classifiers and metrics score how fragile LSB, Unicode, and watermark schemes are under attack.
- Security
- Measurement assumes attackers and transforms will stress detectors.
- Python
- Fragility measurements and toolkit scripts are Python over held-out stego cases.
- Research
- The thesis is measurement of fragility, not a claim of universal detection.
- CLI
- Operators run fragility measurements from the CLI over held-out cases.
- Data visualization
- Steganalysis toolkit turns structure into diagrams or charts so patterns stay inspectable.
- Agentic
- Agents may help run measurement batches; claims stay tied to held-out results.
- Steganalysis
- The toolkit’s product is comparable fragility measurement for hidden-channel schemes.
- PyTorch
- Optional CNN detector tier uses PyTorch/torchvision backbones for steganalysis features.
Portfolio graph
Maps how the work here connects. Open the Portfolio graph →
- Professional
- Personal
- Skill
- Writing