Contents
Anonymise a photo before posting, not after, because posting is the one step in the process you cannot reverse. Uploading hands the original file to a platform you do not control, publishes a re-encoded copy to systems that index and re-share it, and exposes the picture to face search, reverse image search and geolocation the moment it is public. Editing the file itself is a separate task; this article is about the distinct risks created by the act of posting, which are what make “before” the operative word.
The platform keeps the original and re-encodes the copy
Most platforms strip EXIF metadata on upload, which is routinely mistaken for protection. It is not: they re-encode the visible image rather than rebuild it, and many retain the original file you sent on their own servers. That a metadata wipe leaves the identifying content of the image untouched is the whole subject of can you be identified after removing EXIF data. Two copies you cannot edit therefore exist, the public re-encoded one and the platform’s stored original, and the metadata removal you can see applies only to the version shown to other users, not to the company holding your upload. Neither copy is within your reach to change: you can edit your own file endlessly, but the platform can re-share, cache or hand on the original it stored, regardless of what you later delete from your account.
Re-encoding does not touch the identifying signals
The re-encode a platform applies is lossy compression, and the signals that identify you survive it. The camera’s sensor fingerprint is robust to exactly this kind of processing: Goljan and Fridrich report it “survives a wide range of common image processing operations, including lossy compression, filtering, and gamma adjustment” (Proc. SPIE 6819, 2008), so the device trace can outlast the upload, though on recent computational-photography phones its reliability is itself contested (Iuliani, Fontani, Piva, IEEE Access 2021); whether that fingerprint can trace a photo on its own, once the metadata is already gone, is taken up in can a photo be traced without metadata. The face passes through untouched, where recognition networks such as ArcFace can match it (Deng, Guo, Xue, Zafeiriou, CVPR 2019), with the usual caveat that low-quality faces are far harder, since “Recognition in low quality face datasets is challenging because facial attributes are obscured and degraded” (Kim, Jain, Liu, AdaFace, CVPR 2022). If the plan was to hide a face by blurring it, note that blur and pixelation are only partially reversible (Todt, Hanisch, Strufe, PoPETs 2024), so the removal has to be done properly, on your own machine, before the file ever reaches the platform. Even a faceless photo can leak: Tkachenko and Jedidi found 82 of 349 attributes predictable better than random from a facial image, mostly demographic, concluding that facial analysis can “strip away privacy” (Scientific Reports 2023). None of these signals is added by the upload. They are already in the file you chose to post, which is exactly why the only reliable place to remove them is before it leaves your device.
Publishing exposes the photo to search the instant it is live
A posted photo is a photo indexers can reach. Image-geolocation models place a scene from its pixels alone, with PIGEON reaching 40.4% of its predictions within 25 km of the true location (Haas et al., CVPR 2024) and GeoCLIP localising images worldwide against a learned gallery of coordinates (Vivanco Cepeda, Nayak, Shah, NeurIPS 2023). Reverse image search adds another route, and here a common assumption fails: re-encoding or resizing does not reliably defeat it, because a perceptual hash of the TinEye kind is documented invariant to recompression and rescaling, and is weakened mainly by heavy cropping and mirroring. Face-search services such as PimEyes and Clearview match a face across scraped web images rather than matching your file, and are believed to run on deep recognition embeddings of the ArcFace family, though no vendor discloses which. Editing or re-saving the image after it is live does not quietly unpublish it either, because a copy that has already been scraped keeps matching, so you cannot edit your way out once it is public. Whatever you did not remove before posting becomes searchable against the whole indexed web, not merely visible to your intended audience.
”Before” is the only point of control
Because the upload is irreversible and the platform keeps the original, every reduction has to happen on your own machine first: removing rather than blurring the face and identifying content, cropping the scene, and weakening the sensor fingerprint by downscaling and re-encoding, before the image leaves your device, then reverse-searching the edited file yourself to confirm it does not already match an indexed copy. The platform will not anonymise the photo for you, and once it is posted you cannot withdraw it from the caches, scrapers and search indices that will already have seen it. Doing that editing, channel by channel, is the separate task this article stops short of: how to make a photo untraceable.
Sources
- Goljan, Fridrich (2008). Camera Identification from Cropped and Scaled Images. Proc. SPIE 6819, Security, Forensics, Steganography, and Watermarking of Multimedia Contents X.
- Iuliani, Fontani, Piva (2021). A leak in PRNU based source identification. Questioning fingerprint uniqueness. IEEE Access 9.
- Deng, Guo, Xue, Zafeiriou (2019). ArcFace: Additive Angular Margin Loss for Deep Face Recognition. CVPR 2019.
- Kim, Jain, Liu (2022). AdaFace: Quality Adaptive Margin for Face Recognition. CVPR 2022.
- Tkachenko, Jedidi (2023). A megastudy on the predictability of personal information from facial images. Scientific Reports 13:21073.
- Todt, Hanisch, Strufe (2024). Fantômas: Understanding Face Anonymization Reversibility. Proc. Privacy Enhancing Technologies (PoPETs) 2024.
- Haas, Skreta, Alberti, Finn (2024). PIGEON: Predicting Image Geolocations. CVPR 2024:12893-12902.
- Vivanco Cepeda, Nayak, Shah (2023). GeoCLIP: Clip-Inspired Alignment between Locations and Images for Effective Worldwide Geo-localization. NeurIPS 2023.