5

I accept that to some degree this will depend on the software/exact processes used, but I'd like a general idea as to how "risky" it is, and get a better understanding of how biometrics simplifies matching elements.

I wish to maintain a low profile online, but there are communities where it would be useful for people to have a picture if me. My concern is that photo being used to identify me in "real" photos to track me and link my online profile with meat space friends via facial recognition.

I postulate that using a caricature would provide some protection.

Am I correct in assuming that a hand drawn caricature will not leave meaningful markers for facial recognition software, even if they are highly recogniseable to people who subsequently meet me? How much more recogniseable would a photo turned into a caricature digitally be?

Are there key size guidelines to consider for ? (I assume a 32x32 pixel avatar will convey less identifiable info then a larger pic, but is there any "safe" size for an avatar/low res photo - or - I guess another way to ask is how many pixels of a face are required for AI software to get a statistically useful - for unique identification purposes ?

Similarly, can I assume that 16 colors will yield up to 4 times as much data as black and white - and is there much difference between 256 and 64k color?

davidgo
  • 593
  • 5
  • 11

1 Answers1

8

Caricatures and facial recognition software:

Your assumption that "a hand drawn caricature will not leave meaningful markers for facial recognition software" is being studied. See these results from Abaci and Akgul in "Matching facial attributes in caricature-photograph pairs":

Our facial attribute based matching approach defines 32 qualitative features (i.e., gender, hair color, face shape, etc.) for identifying a subject. Then these qualitative features are used to match a caricature to a photograph. In our experiments, we use genetic algorithms and achieve 69% recognition rate at 10% false acceptance rate.

I discovered a second study by Klare, Bucak, and Jain, titled "Towards Automated Caricature Recognition", and they found:

Using only qualitative features, the matching accuracy (at FAR=10.0%) was improved to nearly 57% (using Logistic Regression+NNMKL+SVM). While the image descriptor-based method performed poorly with respect to the qualitative features, it proved valuable when added to the fusion process: Logistic Regression+NNMKL+SVM+LBP with LDA had an accuracy of 61.9%.

Note: I wouldn't be able to explain this because I myself don't understand how they got their results, but I've shared their findings since they are authority figures.

So, I suppose if someone really wanted to find out who you were, and they had the wherewithal, and a bit of luck, they could conceivably find your identity based on an uploaded caricature, but it seems improbable. Moreover, even if they had the desire to, they'd have to jump through some challenging hoops, unless you crossed paths with a researcher in the area.

Caricatures and human recognition:

I'll add just a bit on your comment that caricatures will make you easier to identify. I was surprised to read this bit of research:

Although caricatures are often gross distortions of faces, they frequently appear to be super-portraits capable of eliciting recognition better than veridical depictions. This may occur because faces are encoded as distinctive feature deviations from a prototype. The exaggeration of these deviations in a caricature may enhance recognition because it emphasizes the features of the face that are encoded.

Crazy, huh?