Over the past month and a half (from the beginning of August 2020), quite a few publications / platforms and resources have already talked / wrote about the Fawkes Algorithm: https://sandlab.cs.uchicago.edu/fawkes/#press .
Among which are Habr , The New York Times , The Verge , etc.
Indeed, in our time (quiet 2020) , the topic
Researchers from the University of Chicago have come up with a cloaking algorithm to protect against facial recognition. We posted the sources on github: https://github.com/Shawn-Shan/fawkes .
In August, I read about this tool (Fawkes Algorithm). And I decided to replace my photos on social networks and on all Internet resources where my real photos are.
But, for starters, I decided to try this tool on my photos:
- does fawkes start;
- check with your own eyes the result of fawkes work;
- check on the declared resources (with the same tools as the author of fawkes tested his algorithm).
Fortunately, https://github.com/Shawn-Shan/fawkes has a fairly detailed and simple instruction on how to work with fawkes.
The creator of fawkes claims that the algorithm protects against:
- Microsoft Azure Face API,
- Amazon Rekognition Face Verification,
- Face ++ Face Search API.
This list is indicated in the "Technical Paper" :
On a personal server, compiled from sources: git clone; pip3 install fawkes. It was not easy, but very simple.
I uploaded my photo to the server "r1.jpg". And according to the instructions, I processed this photo using fawkes.
At the exit I got the second photo: "r1_min_cloaked.png". Hooray, I got a cloak-down photo. Opened the photo "r1_min_cloaked.png" - see it with your own eyes. The changes are noticeable, but not critical. Around the eyes, bridge of the nose and nose - there are not significant blackouts.
After that I decided to check the result ("r1_min_cloaked.png") on the services "Microsoft Azure Face API", "Amazon Rekognition Face Verification" and "Face ++ Face Search API".
Result:
As you can see, the Microsoft Azure Face API neural network showed that the original photo (on the left in the screenshot) and the photo after processing (on the right in the screenshot) are one person. Similar figures were shown for the rest of the verification tool: the neural networks "Amazon Rekognition Face Verification" and "Face ++ Face Search API".
The same is with the "protection" of the photo of other people / persons:
That is, as of mid-September "Fawkes" is not working. Perhaps, of course, Fawkes was operating in August 2020. But in September 2020, it no longer works.
A week ago, I wrote a letter to the Fawkes developer and his Fawkes team, asking them to help confirm the algorithm's work. But I have not received a reply letter yet.
At the moment I have not been able to confirm the work of "Fawkes".