Comment by Lionsion | Hacker News Reader

Comment by Lionsion | Hacker News Reader

I still have a Facebook account (which I'm slowly backing away from), here's how I (hopefully) poisoned their facial recognition data.

My profile photos have always been untagged group photos or me in a Halloween costume, etc.

Upload stock photos:

1. Go to https://www.shutterstock.com/search?searchterm=person+model&... Play with the search terms to get people sort of like you.

2. Download a few dozen images (or more). The site gives you the options of finding other photos of the same model. I figure it's probably good to have the poison data be somewhat self-consistent.

3. Upload them to Facebook and tag yourself and your friends in them. I kept them in a non-public album, so I wouldn't spam my feed with this stuff. I suppose I could make them public later, once they're a little old.

4. I've also done this on my girlfriend's account, as well as a couple close friends.

Reverse tag:

When I used to upload and tag real group photos, I'd swap the tagging (e.g. tag a friend as myself and myself as my friend).

I also try to keep the volume of poison photos far larger than real ones. I don't personally upload and tag any photos of myself anymore.

I've always kept the facial recognition stuff like this turned off, but I don't trust Facebook to not to reverse the setting just because it wants to. Eventually I want to delete my account, but I can't just yet due to event invites.

You're likely just putting a lot of effort into helping them test an edge case where tagged photos really don't match. You're probably doing them more of a service than all the normal account users.

I'm not so sure. If that was really a case they cared about, it would be easy for them to generate their test cases themselves.

Seems like a lot of effort to correctly identify a microscopic portion of their users.

I try to equally tag my face and my wife's face (50/50).

That way (even though she's still tied to me), their system can't determine which is hers vs mine - thus it probably has some really funky results. I no longer appear in auto-tags for reference.

Why are you even using Facebook at this point? Your putting this much effort into “confusing” Facebook yet you still share images of you and your wife. I don’t understand.

It's still an easy way to share images of oneself and ones's spouse. The people citilife is interested in sharing those photos with probably are friends with both citilife & citilife's wife, so they come through fine.

Also, leaving Facebook doesn't do much to remove pictures of you. On my feed, most people are posting pictures of their family and friends (& pets & sometimes food) moreso than of themselves. To get your face off of facebook you need to get your friends & family off of facebook. (& your pets since evidently they can have accounts and post pictures).

Might want to add a step where you make sure to strip any information that allows them to find the source, they might look at EXIF data or even do a reverse image lookup to identify false positives.

Disclaimer: I have almost no understanding of how machine learning works under the hood.

Couldn't this potentially assist in their facial recognition software improving?

My understanding is the current crop of machine learning algorithms depend heavily on good training data ("gold data"). If they're training a model to recognize my face, they need accurate training data for my face.

Reply on news.ycombinator.com