GitHub ↩
Privacy tooling · Windows & Linux

See what your photos leak. Then remove it.

Stripping EXIF is not enough. A JPEG can carry a second complete copy of the image, a PNG hides text in chunks, a video keeps GPS in its container, and the address in the caption is plain text. photoscrub finds all of it, then writes clean copies — passive, local, no upload, no account.

Source & releases ↩ Full catalog
12
classes of finding, incl. free-text PII
0
bytes uploaded anywhere
1.8 KB
offline risk classifier
97.4%
accuracy on held-out metadata profiles
What it finds

The leaks people forget about.

HEIC/HEIF from iPhones and RAW files (DNG, CR2, NEF, ARW, ORF, RW2) are read too.

GPS coordinates — HIGH

Exact position to six decimal places, not a rounded city.

Embedded thumbnail — HIGH

A JPEG can carry a second, complete copy of the original as a 160×120 preview. Removing EXIF does not remove it.

Body serial number — HIGH

Identifies the device and, through purchase history, you.

XMP & IPTC — HIGH

City, country, GPS and byline live in a second place. Cleaning EXIF alone is not enough.

PNG text chunks — MED

tEXt, iTXt, zTXt, tIME and eXIf carry author, description and comments.

Data appended after the image — MED

Invisible in any preview, readable in a hex editor.

GPS in video — HIGH

MP4/MOV has no EXIF; the location sits in the udta atom of the container.

PII in free text — MED

Addresses, postcodes and phone numbers in a caption: the leak nobody looks for, because it is not a tag.

ZIP internals — LOW

.DS_Store, Thumbs.db and AppleDouble sidecars carry usernames and full paths.

Use

Report, scrub or pipe it.

Originals are never modified, and --scrub refuses to write into the source folder.

photoscrub PATH # report only
photoscrub PATH --scrub OUT # write metadata-free copies to OUT
photoscrub PATH --json # machine readable
The model

A classifier you can read.

Publish risk is scored from 15 features by logistic regression trained in-house (tools/train_model.py) on 8 000 synthetic metadata profiles. It is not an LLM: 1.8 KB, runs offline, and every decision comes from coefficients you can read in model.json. It is also how we found our own bug — it once called a PNG whose description contained a home address “safe”, which is why has_free_text_pii exists.

Licence keys are Ed25519-signed payloads bound to one machine, and activation works offline; private keys never leave the issuing machine.