ConsentGuard

A local tool that erases private content from a photo before you share it.

consent-guard project page · screenshots, what it catches, and how to run it.

Drop in a photo or screenshot. It finds faces, number plates, text, handwriting, QR codes and barcodes, covers them with solid black, strips the GPS/camera data, and gives you a clean copy. Nothing leaves your machine.

It will miss things. Detection is automatic but never complete, so look at the result before you share it, and fix anything it missed with the brush. The app labels the content types it is bad at instead of pretending otherwise.

Run it

powershell -ExecutionPolicy Bypass -File main_project\scripts\stage_05_review_export\start_consentguard.ps1

That loads the detectors (about a minute) and opens http://127.0.0.1:7860. For a desktop shortcut, run main_project\scripts\stage_05_review_export\create_desktop_shortcut.ps1 once.

First-time setup, if the Python environment does not exist yet:

powershell -ExecutionPolicy Bypass -File main_project\scripts\stage_02_baseline_model\setup_environment.ps1

How it works

photo  ->  normalize (fix orientation, drop metadata, isolated session)
       ->  detectors: 9-class Mask R-CNN, face, plate, text, QR/barcode, EXIF
       ->  second pass over overlapping tiles, so small regions are found
       ->  fuse: per-class thresholds, area sanity caps, merge overlaps
       ->  solid black fill of the mask shape, grown a few pixels
       ->  encode a fresh file, re-open it, re-scan it for leftovers
       ->  download (same format as the input)

You either press Erase and save and get the file, or open the review screen first to paint over misses and un-erase anything covered by mistake. Working copies are deleted when you close the page; your original is never modified.

What it catches

Measured on 120 validation photos the models never trained on: 83% of all labelled private pixels get covered (95% CI 78–87%).

Expect erasure even on photos with nothing private in them. The tool erases all text by design, and on text-heavy photos — labels, adverts, documents, signs — that is most of the picture. On 30 validation photos with no private visual content, every one had something erased, 45% of each photo on average. A diagnostic on six of them traced ~40 points of that to text (the detector was right: it was real text) and ~6 points to stray plate boxes on photos with no plates. Switch text off under What to erase when you only want faces and plates gone. Full report: reports/PERSONAL_PROFILE_SCORECARD_2026-09-18.md.

Content	Private pixels covered	Whole regions caught	Verdict
Faces	99.8%	95%	Good
Whole people	94.8%	83%	Good, and off by default — it erases most of a photo
Number plates	95.4%	56%	Catches many; small and distant ones are missed
Nudity	93.4%	33%	Check manually
Signatures	65.8%	23%	Weak
Medicine and documents	52.1%	7%	Weak
Handwriting	48.5%	50%	Weak
Other details (fingerprints, visible disability)	—	—	Best effort only. No working detector for either; whatever the broad model flags is still erased, but they are not offered as protected categories
QR codes and barcodes	—	—	Reliable: a decoder, not a model
GPS/camera metadata	—	—	Always removed; the file is written from scratch

"Whole regions caught" counts regions where a mask of that type covered at least half the region, which is the number that matters when one small plate is the thing you needed hidden. Those figures are from the single-pass configuration.

The app also runs a tiled second pass, which is what rescues small regions. Measured on 20 real road photos holding 40 annotated number plates: it covers 33 of 40 plates (82.5%) against 24 of 40 (60%) single-pass, for 1.7 points more of the image erased.

Reproduce them:

.\.venv\Scripts\python.exe main_project\scripts\stage_06_evaluation_release\evaluate_fused_validation.py `
  --config main_project\configs\stage_02_baseline_model\train_maskrcnn_moderate_v2_negatives_10ep.yaml `
  --checkpoint artifacts\checkpoints\maskrcnn_moderate_v2_negatives_10ep\last.pt `
  --threshold-profile main_project\configs\stage_04_fusion_calibration\threshold_profile_personal_max_coverage.yaml `
  --device cuda --max-images 120 --output reports\personal_profile_validation_scorecard.json

measure_mask_coverage.py answers the narrower question of whether annotated regions actually end up under the burnt-in mask, by driving the running app.

What it will not do

• Guarantee anything. It reduces risk; it does not certify a photo as safe.
• Decide whether anyone consented. You decide what to keep.
• Work on video, or help with a photo you already shared.
• Send anything anywhere. There is no server and no telemetry.

Checks

powershell -ExecutionPolicy Bypass -File main_project\scripts\ci_local.ps1

Python tests, the frozen baseline hash check, the frontend build, frontend unit tests, and the browser end-to-end flow. Add -IncludeModels for a real one-photo run through every detector. There is no hosted CI: the tests need the local GPU, the checkpoints and the dataset records, none of which leave this machine.

Layout

• main_project/ — the code, in six numbered stages (data, baseline model, specialist detectors, fusion, review/export, evaluation and release).
• main_project/scripts/stage_05_review_export/ — the app and its launcher.
• main_project/configs/stage_04_fusion_calibration/ — threshold profiles. The personal profile is what the app uses; the validation-calibrated one is stricter and precision-oriented.
• reports/ — measured evidence, one file per run.
• docs/archive/ — earlier research reports, failure registers and download guides, with an index explaining what is still true.
• research/ — training and Kaggle machinery. Not needed for the app.

Datasets, checkpoints and outputs stay out of version control and must not be redistributed without their original licences.

Honest status

The detectors are single-seed research checkpoints. The threshold profile is tuned, not certified: release_ready: false is in that file on purpose, and the app shows it as a warning rather than blocking the download. The strict release gates in main_project/stage_06_evaluation_release/ (95% recall in two domains, three seeds, independent attackers, a 2,000-photo target set) are not met and are not being pursued. A person reviewing the result is the safety net instead.

The research background, including why the project is scoped this way, is in ConsentGuard_Final_Research_Design.md.