Contents
More than you would expect, and most of it before you type a single word of caption. One photograph can disclose the device that took it, where and roughly when it was taken, who is in it, and a good deal about what those people are like, and most of that survives even if you strip the file’s metadata first. This is a field guide to your own image’s disclosure surface, working from the part that is easy to clear to the parts that are not.
The layer you can clear
The removable layer is the metadata: EXIF, XMP and IPTC blocks holding the camera model, the timestamp and often precise GPS coordinates. It is structured text stored alongside the picture, so a remover or a simple re-save strips it without changing any pixel. Clearing it is worthwhile, and it is also the smallest part of what your photo reveals. Everything below is read from the picture itself and survives the wipe, which is why metadata removal alone does not anonymise a photo (does removing metadata make a photo anonymous).
Your device
Beneath the image is the camera’s own noise signature, its photo-response non-uniformity, or PRNU, a per-device pattern introduced for camera source identification by Lukáš, Fridrich and Goljan (IEEE TIFS 2006). It comes from minute manufacturing differences between individual sensors, so it lives in the pixels rather than in any tag, and it is robust: Goljan and Fridrich report that “The fingerprint survives a wide range of common image processing operations, including lossy compression, filtering, and gamma adjustment” (Proc. SPIE 6819, 2008). The fingerprint is conditional in practice, it needs the candidate device and clean reference shots to match against, and its uniqueness is contested on modern phones (Iuliani, Fontani, Piva, IEEE Access 2021), but the signal itself does not leave when your tags do. The mechanics are unpacked in can a photo be traced without metadata.
Your location
The scene gives away where you were, with no GPS tag required. Neural geolocation models place a photo on a map from its visual content alone: PlaNet reached “superhuman levels of accuracy in some cases” (Weyand, Kostrikov, Philbin, ECCV 2016), and later systems are more precise. PIGEON lands 40.4% of its predictions within 25 km of the true location, and separately reports 92.0% country accuracy with a median error of 44.4 km (Haas et al., CVPR 2024); the 25 km figure is city-scale and the 92.0% figure is country classification, two different measures. GeoCLIP matches an image against a learned, continuous map of coordinates rather than a fixed list of places (Vivanco Cepeda et al., NeurIPS 2023). A storefront, a street sign, a mountain profile or even a reflection can narrow a place your deleted coordinates were meant to protect.
Your identity
Then there is you. A clear face is close to fully identifiable to modern recognition networks: ArcFace reaches about 99.8% on the LFW benchmark (Deng et al., CVPR 2019). That headline needs a caveat beside it, since Kim and colleagues note that “Recognition in low quality face datasets is challenging because facial attributes are obscured and degraded” (AdaFace, CVPR 2022), so small or low-quality faces are far less reliable. Even without a clean match, an image discloses attributes. Tkachenko and Jedidi ran a megastudy over 2,646 facial images of 969 individuals and 349 binary attributes, and found 82 of them, about 23%, predictable better than random from the picture alone (Scientific Reports 2023). That readable signal is mostly demographic, and the authors name “smartphone camera artifacts, BMI, skin properties, and facial hair” as top candidate non-demographic signals, not personality or intelligence. The apparent is not the true, yet the authors still describe the net effect plainly, that facial analysis retains “the ability … to strip away privacy.”
It is tempting to assume that blurring the face closes this channel. Todt, Hanisch and Strufe found Gaussian blur and pixelation to be partially reversible, because they “only generalize the information … by averaging it,” leaving identity obfuscated and not removed (Fantômas, PoPETs 2024). Stronger, GAN-based anonymisers were much harder to reverse in the same study, so the choice of tool is what makes the difference.
Reading your own photo before you share it
Add these layers up and the lesson is not that every photo is a disaster, but that the clearable tags are the smallest piece of it. What a given image reveals depends on the image: one with no face, no distinctive background and heavy compression discloses little, while a sharp, geolocatable selfie discloses almost everything regardless of its tags. A useful check before posting is four short questions: are the tags gone, is the scene identifying, is the face identifying, and would the device fingerprint matter for whoever might examine the file. When you decide to reduce that surface, the practical routes are how to make a photo untraceable and how to anonymise a photo before posting.
Sources
- Lukáš, Fridrich, Goljan (2006). Digital Camera Identification from Sensor Pattern Noise. IEEE TIFS 1(2):205-214.
- 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.
- Weyand, Kostrikov, Philbin (2016). PlaNet: Photo Geolocation with Convolutional Neural Networks. ECCV 2016.
- 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.
- Deng, Guo, Xue, Zafeiriou (2019). ArcFace: Additive Angular Margin Loss for Deep Face Recognition. CVPR 2019:4690-4699.
- 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(4).