Popular dating service Bumble has just announced the release of an open source version of its Private Detector A.I. tool, which was created to shield their community from unwanted lewd images. Cyber-flashing is an unfortunate and common side effect of the growth in personal relationship services, and service providers spend millions of dollars trying to combat the problem.
Now, with this announcement, the company hopes that the wider developer community will help to enhance and evolve the tool to better protect community users across the world.
Private Detector works by automatically blurring a potentially nude image shared in a chat on Bumble. The platform notifies the user with a warning, and it’s then up to them to decide whether to view or block the flagged image.
Bumble’s developers trained the AI using very high volume data sets, with any negative samples (i.e. the ones not containing any lewd content) carefully flagged. They did this to better identify edge cases and recognize parts of the human body, such as legs and arms, to avoid falsely flagging them as abusive.
Developers initially released Private Detector in 2019 as part of the company’s long-term privacy and security program. The company has stated that the tool achieves a performance accuracy of over 98%, with no trade-offs in precision. This is even when analyzing performance in different conditions, both offline and online.
Release Of A.I. Source Code To Benefit All
Private Detector’s latest version is being released under the Apache License, to encourage adoption of the technology as a general standard for blurring lewd images. The release will also be accompanied by a white paper describing the technology behind Private Detector, and the source code of the open-source version will now be available on the GitHub developer platform, for anyone to download and develop on.
The code provides the methodology used to train the AI engine behind Private Detector. The company is also releasing a ready-to-use model, to allow developers to deploy the technology immediately.
Finally, developers can use a machine learning checkpoint model to help them fine-tune the engine with additional images to improve performance. In this way, the company hopes to incentivize the developer community to improve the accuracy of the results.
This announcement is a welcome development in the fight against online harassment and abuse. By releasing the technology to the open source community, there’s every chance that many more companies who face the same policing difficulties will be able to shortcut a path to an effective solution.
This will be especially important for startups without the resources to spend time and money building their own A.I. tool to combat the cyber-flashing problem. The more companies there are which can deploy this kind of tool, the better. Machine learning is not cheap, as it requires a careful development cycle of a neural network, and thousands of hours of work validating the image data and weeding out false positives.