Someone scraped 40,000 Tinder selfies to make a facial dataset for AI experiments

Tinder users have many motives for uploading their likeness to the dating app. But contributing a facial biometric to a downloadable data set for training convolutional neural networks probably wasn’t top of their list when they signed up to swipe.

A user of Kaggle, a platform for machine learning and data science competitions which was recently acquired by Google, has uploaded a facial data set he says was created by exploiting Tinder’s API to scrape 40,000 profile photos from Bay Area users of the dating app — 20,000 apiece from profiles of each gender.

The data set, called People of Tinder, consists of six downloadable zip files, with four containing around 10,000 profile photos each and two files with sample sets of around 500 images per gender.

Some users have had multiple photos scraped from their profiles, so there is likely a lot fewer than 40,000 Tinder users represented here.

The creator of the data set, Stuart Colianni, has released it under a CC0: Public Domain License and also uploaded his scraper script to GitHub.

He describes it as a “simple script to scrape Tinder profile photos for the purpose of creating a facial dataset,” saying his inspiration for creating the scraper was disappointment working with other facial data sets. He also describes Tinder as offering “near unlimited access to create a facial data set” and says scraping the app offers “an extremely efficient way to collect such data.”

“I have often been disappointed,” he writes of other facial data sets. “The datasets tend to be extremely strict in their structure, and are usually too small. Tinder gives you access to thousands of people within miles of you. Why not leverage Tinder to build a better, larger facial dataset?”

Why not — except, perhaps, the privacy of thousands of individuals whose facial biometrics you’re dumping online in a mass repository for public repurposing, entirely without their say-so.

Glancing through a few of the images from one of the downloadable files they certainly look like the sort of quasi-intimate photos people use for profiles on Tinder (or indeed, for other online social apps) — with a mix of selfies, friend group shots and random stuff like photos of cute animals or memes. It’s by no means a flawless data set if it’s just faces you’re looking for.

Reverse image searching several of the photos mostly drew blanks for exact matches online, so it appears that many of the photos have not been uploaded to the open web — though I was able to identify one profile image via this method: a student at San Jose State University, who had used the same image for another social profile.

She confirmed to TechCrunch she had joined Tinder “briefly a while back,” and said she doesn’t really use it anymore. Asked if she was happy at her data being repurposed to feed an AI model she told us: “I don’t like the idea of people using my pictures for some sad ‘researches.’ ” She preferred not to be identified for this article.

Colianni writes that he plans to use the data set with Google’s TensorFlow’s Inception (for training image classifiers) to try to create a convolutional neural network capable of distinguishing between men and women. (I just hope he strips out all the pet shots first or he’ll find this task an uphill struggle.)

The data set, which was uploaded to Kaggle three days ago (minus the sample files), has been downloaded more than 300 times at this point — and there’s obviously no way to know what additional uses it might be being put to.

Developers have done all sorts of weird, wacky and creepy things playing around with Tinder’s (ostensibly) private API over the years, including hacking it to automatically like every potential date to save on thumb-swipes; offering a paid look-up service for people to check up on whether a person they know is using Tinder; and even building a catfishing system to snare horny bros and make them unwittingly flirt with each other.

So you could argue that anyone creating a profile on Tinder should be prepared for their data to leech outside the community’s porous walls in various different ways — be it as a single screenshot, or via one of the aforementioned API hacks.

But the mass harvesting of thousands of Tinder profile photos to act as fodder for feeding AI models does feel like another line is being crossed. In the scramble for big data sets to fuel AI utility, clearly very little is sacred.

It’s also worth noting that in agreeing to the company’s T&Cs Tinder users grant it a “worldwide, transferable, sub-licensable, royalty-free, right and license to host, store, use, copy, display, reproduce, adapt, edit, publish, modify and distribute” their content — though it’s less clear whether that would apply in this case where a third-party developer is scraping Tinder data and releasing it under a public domain license.

At the time of writing Tinder had not responded to a request for comment on this use of its API. But since Tinder makes its rights to your content transferable, it’s entirely possible even this large-scale repurposing of the data falls within the scope of its T&Cs, assuming it sanctioned Colianni’s use of its API.

Update: A Tinder spokesperson has now provided the following statement:

We take the security and privacy of our users seriously and have tools and systems in place to uphold the integrity of our platform. It’s important to note that Tinder is free and used in more than 190 countries, and the images that we serve are profile images, which are available to anyone swiping on the app. We are always working to improve the Tinder experience and continue to implement measures against the automated use of our API, which includes steps to deter and prevent scraping.

This person has violated our terms of service (Sec. 11) and we are taking appropriate action and investigating further.