Blurring a face can make someone harder to recognize, but it does not make a photograph reliably anonymous. Pixelation can retain useful image information; a solid cover can hide facial pixels while leaving recognizable hair, clothing or context. If a person's identity must remain private, removing them from the shared image or choosing another photograph is a better starting point than increasing a blur slider.
The useful question is not simply “Can I still see the face?” It is “Could the intended audience identify this person from the complete image, its accompanying text or other available photos?” This guide separates what research actually tested from a practical check you can do before sharing.
Three different meanings of “hidden”
A photograph can have reduced facial detail, such as a blurred portrait; covered facial pixels, such as a fully opaque patch in an exported image; or an unidentified person. These are different outcomes. The first two describe an edit. The third depends on what remains visible, who sees it and what else they know.
An emoji over the eyes may satisfy a visual style without covering the rest of the face. A large opaque patch may cover the entire head while a name badge identifies the person below it. Neither example requires anyone to recover the original face. Keep this distinction in mind when reading claims that an edit can be “reversed.”
What the evidence says about blur and pixelation
In the 2016 research preprint Defeating Image Obfuscation with Deep Learning, McPherson, Shokri and Shmatikov tested pixelated and blurred images. Models classified identities despite reduced detail.
Candidate identities were represented in training, and models learned from correspondingly obfuscated images. This does not demonstrate that any blurred stranger can be named or exact original facial pixels reconstructed. These historical experiments do not evaluate every current app. They show why looking unreadable is insufficient evidence that an image retains no information useful for recognition.
The work establishes no universally safe blur radius or mosaic size. Resolution, processing and reference material differ. Judging stronger blur by eye cannot settle those variables.
A covered face can still belong to a recognizable person
Oh and colleagues' ECCV 2016 study examined person recognition in social photo albums using the PIPA dataset. With labeled reference photographs and recurring identities, information outside the head and relationships between photos could support recognition even after the head was obscured. Results depended on whether photographs came from the same event or different events.
That does not show that a solid cover reveals the pixels beneath it. It shows that concealing the head and concealing identity are separate problems. Someone wearing the same distinctive jacket in a labeled group photograph and an obscured photograph may be recognizable from the connection between them.
People can also recognize familiar appearances without a detailed face. In a 2022 pilot study by Preston and colleagues, some celebrity photographs remained recognizable when a black box covered most facial features, leaving hair and neck exposed. The first survey involved 16 university students or staff; the images were selected for celebrity familiarity. This small, structured experiment does not provide a recognition probability for ordinary strangers or test emoji stickers. It does explain why “my friend would know who this is” deserves attention even when an unfamiliar viewer cannot tell.
Choose the edit according to what you need to protect
The following table compares visible changes, not certified levels of anonymity. Every option still needs a check of the complete shared file and its context.
| Option | What it changes | What still needs checking |
|---|---|---|
| Blur or pixelation | Reduces detail while preserving some image structure. | Remaining facial information, exposed features and every cue outside the edited area. Do not treat a strong-looking effect as a guarantee. |
| Fully opaque coverage | Replaces the covered pixels in a properly exported composite. | Complete coverage, opacity, hair or head outline, clothing, reflections and whether you are sending the exported copy rather than an editable project. |
| Crop the person entirely out | Removes the person from the visible frame of the sharing copy. | Reflections, other views, captions and whether the original or an additional image is also attached. |
| Choose another image or do not share | Avoids publishing this depiction. | Whether the replacement image or accompanying text discloses the same person. Use this option when remaining uncertainty is unacceptable. |
A sticker is not automatically an opaque redaction: its artwork can contain transparent areas, and its position may leave identifying features exposed. Likewise, a crop inside an editing project is not the same thing as checking the exported image you intend to send.
A seven-step check of the actual sharing copy
- Decide whether the person needs to appear. State the purpose and audience first. A public event recap may work with a detail shot of the venue instead of a recognizable bystander. Ask whether the person is comfortable with the planned sharing.
- Check every person, not just the main subject. Scan the edges, background, mirrors, windows and any smaller images within the picture. A face hidden in the center does not protect someone visible elsewhere.
- Choose deliberate coverage or removal. If you keep the person, inspect the complete head and surrounding details rather than covering only the eyes. Check the boundaries at full image size. Do not infer safety from how small the preview looks.
- Read the rest of the frame. Look for visible names, badges, distinctive belongings and links to other images. NIST's review of de-identification explains why multimedia can carry identifying information through several kinds of cues together. Removing one cue does not evaluate the others.
- Export a separate sharing copy. Use a flattened still-image export that incorporates the edits into its visible pixels, such as a new JPEG or PNG. Keep the original and editable project out of the attachments. Do not assume a format name alone proves everything sensitive was removed.
- Reopen the exact exported file. View it in a separate image viewer, inspect the edits and check the actual attachment before sending. Review its filename, caption, tags and additional images too. This catches visible mistakes; it is not a forensic anonymity test.
- Stop if the remaining cues defeat the purpose. If colleagues would recognize the clothes, or the caption names the only person present, choose a different image, remove more of the scene or do not post. A restricted audience can reduce exposure, but recipients may still recognize someone or forward the file.
The EFF's photo-sharing guidance likewise encourages considering permission and identifying details beyond faces. This checklist is a practical review process, not a legal determination that a photograph has been anonymized.
Example: the event photograph with a hidden bystander
Imagine you want to share a photograph of yourself at a workshop. A bystander asks not to be identifiable, so you blur their face. Their badge is still readable, and the next image in the carousel shows them unedited. Increasing the blur on the first image changes neither disclosure.
A useful revision starts with the sharing goal. Crop the bystander entirely out if the picture still works, remove the second image, and check the caption for their name. If a reflection still shows them or the event context identifies them, use another photograph. No recognition service needs to be consulted to notice these problems.
Keep identity, location and appearance checks separate
Visible identity is only one part of photo sharing. A file can reveal location independently of its face edits; use our location metadata and sharing checklist for that separate check. Removing location information does not establish anonymity either.
An attractive-looking edit has another purpose again. Our guides to beauty filters and first impressions and AI headshots and honest representation discuss how presentation changes what a portrait communicates. Neither a convincing generated face nor a favorable appearance assessment establishes privacy.
FaceScore offers photo-based appearance feedback. If you choose to use it, upload a photograph you have permission to use, with other people cropped out. Treat its feedback as information about that image, not a test of whether a person can be identified. When protecting identity is the priority, decide what to share before seeking feedback on how it looks.