Image Forensics Inspector

Image Forensics Inspector (Skill)

This skill exposes the find_image_hide project as a reusable capability for OpenClaw, Hermes, and other agents. Everything runs 100% locally — no network calls, no telemetry, no SaaS dependency.

When to invoke this skill

Invoke when the user/agent task involves any of:

• "Is this image AI-generated / has C2PA / SynthID-like signal?"
• "Is anything hidden inside this PNG/JPG (zip, text, executable, LSB payload)?"
• "Was this image tampered with (copy-move, splicing, ELA bright regions)?"
• "Does this image carry a Shutterstock / Getty / Adobe Firefly watermark?"
• "Reverse-lookup this image against my local reference library (pHash)."
• "Produce a court-grade, byte-reproducible PDF forensic report."
• Batch scanning a folder of images for risk triage (LOW / MEDIUM / HIGH).

Capability inventory

Mapped to source modules so the agent can cite evidence chain:

Prerequisites

1. Python 3.10+.

2. Install runtime deps:

   cd d:\workspace\project\find_image_hide
   python -m venv .venv
   .venv\Scripts\activate
   pip install -r requirements.txt
   

3. Optional soft-dependencies (skill degrades gracefully without them — fields are tagged SKIPPED_NO_LIBRARY / SKIPPED_NO_TOOL):

   • pip install reportlab — needed for the byte-reproducible PDF report.
   • pip install invisible-watermark opencv-python — DwtDct invisible-watermark decode.
   • pip install c2pa-python — pure-Python C2PA verification (or system-wide c2patool).
   • System: tesseract (visible-watermark OCR), binwalk / zsteg / stegoveritas / stegseek (external steg tools).
   • Env var FORENSICS_PHASH_REFERENCE_DIR=<absolute dir> — pHash reverse-lookup library.

Usage modes

Mode A — CLI (recommended for agent automation)

Entry point: analyze_image.py

Single image:

python analyze_image.py --input <path-to-image> --output <out-dir>

Directory (batch):

python analyze_image.py --input <dir> --output <out-dir> --recursive --workers 4

stdout (single-image mode) is a JSON object:

{
  "report": "<abs-path>/report.json",
  "risk_level": "LOW|MEDIUM|HIGH|UNKNOWN",
  "confidence": 0.0
}

Full machine-readable result lives in <out-dir>/report.json (single) or <out-dir>/summary.json + <out-dir>/<slug>__<hash>/report.json (batch). Schema is summarized in README.md.

Mode B — Python API (in-process embedding)

from pathlib import Path
from image_forensics.analyzer import analyze_image
from image_forensics.batch import analyze_directory

report = analyze_image(Path("test.jpg"), Path("./out"))
risk = report["overall"]["risk_level"]

summary = analyze_directory(
    Path("./photos"), Path("./out"),
    recursive=True, workers=4,
)

Both functions write report.json / summary.json to the output dir and return the same dict.

Mode C — Local HTTP API (for non-Python agents like Hermes)

Boot the local server (binds to 127.0.0.1 only):

python webapp.py --host 127.0.0.1 --port 5050

Then call:

Full server impl: webapp.py.

Mode D — One-shot demo

python demo.py            # prepare samples + run
python demo.py --serve    # + start web UI
python demo.py --no-download  # offline-only synthetic samples

See demo.py.

Output contract

report.json schema (full list in README.md L306-L326):

{
  "schema_version": "0.1.0",
  "input": { "file_name": "...", "format": "PNG", "sha256": "...", ... },
  "overall": {
    "risk_level": "LOW|MEDIUM|HIGH|UNKNOWN",
    "confidence": 0.0,
    "summary": "...",
    "module_scores": { "fft": 0.0, "dct": 0.0, "lsb": 0.0, "noise": 0.0, "metadata": 0.0, "provenance": 0.0 }
  },
  "metadata": { ... },
  "ai_provenance": { "status": "NO_PROVENANCE_FOUND | VERIFIED_AI_GENERATED | VERIFIED_AI_EDITED | PROVENANCE_PRESENT_BUT_UNVERIFIED | POSSIBLE_AI_BUT_UNVERIFIED", ... },
  "frequency_analysis": { ... },
  "dct_analysis": { ... },
  "lsb_analysis": { ... },
  "noise_analysis": { ... },
  "evidence_items": [ ... ]
}

Visualizations written alongside report.json:

visualizations/
  spectrum.png, r/g/b_spectrum.png
  dct_mean_heatmap.png, dct_histogram.png
  lsb_r.png, lsb_g.png, lsb_b.png
  residual.png, laplacian.png

Recommended agent recipes

1. Triage a folder for an OpenClaw task — run Mode A in batch, then parse summary.json and surface every entry where risk_level == "HIGH"; deep-dive each by reading its report.json and citing evidence_items[].description.
2. Hermes "is this AI?" check — run Mode A single-image, read report["ai_provenance"]["status"]. Treat VERIFIED_AI_GENERATED / VERIFIED_AI_EDITED as hard evidence; POSSIBLE_AI_BUT_UNVERIFIED as soft signal that needs human review; never claim hard AI judgement based solely on ai_heuristics.
3. CTF / payload hunt — read report["extraction"] and report["external_steganalysis"]; if either reports trailing zip / archive / stegseek hit, extract the file path it emits and follow up.
4. Court-grade evidence — call /api/jobs/.../report.pdf and capture the X-Pdf-Sha256 + X-Source-Sha256 response headers into the case file. Re-rendering the same report.json on any machine must produce the same PDF SHA-256 (see P2.4_ACCEPTANCE.md).

Honesty / limits the agent MUST surface

• This skill cannot prove an image is or is not AI-generated; ai_heuristics is hand-tuned thresholds, not a trained classifier. See the "🟠 C 档" section of README.md.
• Classic χ²/SPA steganalysis is ineffective against modern adaptive steg (HUGO / WOW / S-UNIWARD).
• SynthID has no official local decoder — only Google's API is authoritative.
• LSB confidence is auto-down-weighted on lossy formats (JPEG).
• No watermark is ever removed, no C2PA is ever forged; this is a read-only forensic tool.

Security envelope

• All processing is local; no outbound traffic.
• HTTP server defaults to 127.0.0.1; agents binding 0.0.0.0 accept full responsibility.
• Uploaded files are path-sanitized (.. rejected, extension whitelisted) and confined to analysis_output/<job_id>/_uploaded/.
• Upload size cap defaults to 1 GB (see MAX_UPLOAD_BYTES in webapp.py).