Preventing computer-aided computer-aided pattern matching/"dupe detection" of photos?

[Mike64]

I'm not sure this is the right sub-forum for this, but it seemed like the best fit. Mods should of course feel free to move it And I know this is a vague question and what I'm looking for may be even theoretically impossible, but I don't know enough about the underlying subject(s) to be more specific. But anyway...

Are there recognized techniques (and/or even an existing utility) for altering image files to make computer-aided matching more difficult, if not "impossible", while leaving the "images" in them relatively unscathed to the human eye/brain? The usage I'm thinking of is posting photos of people online or distributing them directly as files, where you want someone/others to be able to definitely see what the subject/person "looks like", though not necessarily in great detail, and aren't concerned about the subjects being recognized "in person", but you do want to prevent them being mass/high-speed scanned for matching photos, or at least make that totally impractical for an amateur with a (even heavy-duty) home PC to do it...

ETA: In case it wasn't clear, I'm not concerned about very sophisticated pattern-matching comparisons (like the sort of thing I imagine that heavy-hitting LE agencies like the FBI have at their disposal), just stuff "nosy people" with technical knowledge might do - things I assume are akin to the way ordinary photo "dupe-checking" programs work, but I don't use them much so I'm not very clear on how broad their comparisons are. Would simply closely cropping photos, or drastically lowering their resolution, do it?

 

[agent00f]

I'm not sure this is the right sub-forum for this, but it seemed like the best fit. Mods should of course feel free to move it And I know this is a vague question and what I'm looking for may be even theoretically impossible, but I don't know enough about the underlying subject(s) to be more specific. But anyway...

Are there recognized techniques (and/or even an existing utility) for altering image files to make computer-aided matching more difficult, if not "impossible", while leaving the "images" in them relatively unscathed to the human eye/brain? The usage I'm thinking of is posting photos of people online or distributing them directly as files, where you want someone/others to be able to definitely see what the subject/person "looks like", though not necessarily in great detail, and aren't concerned about the subjects being recognized "in person", but you do want to prevent them being mass/high-speed scanned for matching photos, or at least make that totally impractical for an amateur with a (even heavy-duty) home PC to do it...

ETA: In case it wasn't clear, I'm not concerned about very sophisticated pattern-matching comparisons (like the sort of thing I imagine that heavy-hitting LE agencies like the FBI have at their disposal), just stuff "nosy people" with technical knowledge might do - things I assume are akin to the way ordinary photo "dupe-checking" programs work, but I don't use them much so I'm not very clear on how broad their comparisons are. Would simply closely cropping photos, or drastically lowering their resolution, do it?

This really depends on how sophisticated class of alg you're trying to fool. If it's a very simple hash or blocked-based whatever (searching for near identical image) then doing much of anything like shifting it over by half a block will do the trick.

But consider for somewhat more advanced algs, which is probably what google/bing type image searches use: http://vision.caltech.edu/malaa/software/research/image-search/, I'm not sure there's a general way to defeat all the common/plausible ones. For example, some build a histogram of colors, so you'll have to shift/alter colors past whatever bin size and whatever other heuristic they're using in their implementation, and that's just one simple-ish method.

 

[Mike64]

This really depends on how sophisticated class of alg you're trying to fool. If it's a very simple hash or blocked-based whatever (searching for near identical image) then doing much of anything like shifting it over by half a block will do the trick.

But consider for somewhat more advanced algs, which is probably what google/bing type image searches use: http://vision.caltech.edu/malaa/software/research/image-search/, I'm not sure there's a general way to defeat all the common/plausible ones. For example, some build a histogram of colors, so you'll have to shift/alter colors past whatever bin size and whatever other heuristic they're using in their implementation, and that's just one simple-ish method.

Thanks, and especially thanks for the link. Between the specific info there, and with that as guide to further research, I should be able to figure out something that gets me close enough to what I'm looking for.

 

[slugg]

Can you be more specific as to what kind of programs you're trying to fool? I don't really know what nosy people do with photos online. But I do know image processing and recognition, so I'll see what I can do once I better understand your problem.