Facebook is trialing a service in Australia where people upload their sexy, embarrassing images to prevent them from being published in the social network. This service uses a technical procedure called “hashing.”
Using hashing algorithms to take a mathematical summary or “fingerprint” of a photo is a cool technique. It can be used to recognize an image that has been seen before without actually viewing the image. This is very useful for detecting duplicates and creating watch lists.
Special “fuzzy” hashing techniques can even recognize a photo when it has been cropped or changed slightly. Microsoft has developed a PhotoDNA technique that is particularly good at this, and it is a key tool used by INTERPOL and other law enforcement agencies for processing child exploitation images. By quickly recognizing images that have been seen before, investigators can focus their efforts on new images that might provide fresh evidence.
Microsoft also offers a cloud service so that online sites can screen images that are uploaded by users.
Now Facebook is starting a trial using the same kind of technology. The idea is that revenge-porn images, usually explicit images being distributed without the subject’s knowledge and consent, can be recognized and blocked before they are put online.
This is an interesting idea, but how is Facebook to know if the image being submitted for blocking is actually a revenge-porn image – maybe someone is instead trying to interfere with their business competitors, for example.
The solution that Facebook describes is to have staff members review each submitted image to make sure that it violates their policies on explicit images. Obviously this is rather privacy invasive, and presents further risks of exposure and embarrassment.
More concerning is that Facebook intends to keep the submitted images for some time (with some form of blurring), instead of immediately deleting them once the hashes are created. This obviously creates further privacy and security risks.
Facebook is trying to tackle a serious problem using effective technical methods, but the devil is going to be in the implementation. It is not clear why the images have to be retained, for example, once the hashes are created. It will be interesting to follow the results of the trial – hopefully Facebook will let us know how it worked out.
Motherboard has provided some more information. Now it appears that the images will not be blurred once they are submitted, but they are also not retained after the hashes are created.
It seems to be rather difficult to get a clear description of how this program will really work and, as I mentioned, the devil is in the details.