0
$\begingroup$

To give some context:

I am looking for a suite of techniques and tools that can theoretically enable me to conduct analysis such as classification on image datasets in a manner in which a naive observer cannot derive "much" information from an image e.g. if the image is of a person/object then the naive observer cannot tell who/what is in the image.

Naively adding noise doesn't seem to protect the content of an image, since a human observer can easily see past the noise.

I'm not aware that a machine learning task can be performed on encrypted images effectively and on naive inspection it shouldn't be possible while still maintaining cryptographic properties.

Is there any work out there on such techniques that maintain privacy of an image while still allowing it to be used for analysis?

Is what I'm after a fool's errand?

$\endgroup$
2
  • $\begingroup$ Do you want to process on encrypted data? $\endgroup$
    – kelalaka
    Jan 8, 2020 at 18:47
  • $\begingroup$ I would like to process on image data that has it privacy enhanced i.e. some privacy mechanism has been applied to the image data such that useful analysis can be done on it (maybe via ML) while still making the content of the image non-trivially processed by a human (imagine you have given your photo to me, if your neighbour can recognize you immediately after applying this privacy mechanism, it probably is not private enough). But a ML algorithm should still be able to determine the picture is of you on a boat. $\endgroup$
    – hdu
    Jan 16, 2020 at 22:31

2 Answers 2

1
$\begingroup$

I think what you are looking for is "Perceptual Hashing".

What I understand from your question is that you need a way for programmatically identifying images even when they have been manipulated beyond human recognition.

Their is a Microsoft project called PhotoDNA. I believe the have a cloud API.

You may also want to look at this Python example

$\endgroup$
1
  • $\begingroup$ I think you are getting me to the right direction. I will look a little deeper into this and mark your answer if I think it answers it! Thanks! $\endgroup$
    – hdu
    Jan 16, 2020 at 22:23
0
$\begingroup$

I don't know if it will give you a practical solution. But some people have tried to use fully homomorphic encryption to do picture analysis:

https://eprint.iacr.org/2017/1114

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.