A New AI-Powered Software
A programmer invented an application that removes clothing from photos of women, making them appear really naked. DeepNude software takes a snapshot of a clothed person and generates a new, naked image of that same person. It only works on photographs of women and swaps garments for nude breasts and a vulva. When Motherboard tried to use a man’s photograph, it substituted his pants with a vulva.
Effectiveness and Concerns
While DeepNude works on photographs of fully clothed women with varied degrees of effectiveness, it tends to perform best on images where the subject is already revealing a lot of flesh. We tested the software on dozens of photos and found that high-definition images from Sports Illustrated Swimsuit issues produced the most convincing results.
Deepfakes have been a source of concern for the media and politics since Motherboard found them in late 2017. But the most damaging application of deepfakes has always been against women: whether to experiment with the technology using photographs without women’s consent, or intentionally disseminating nonconsensual porn on the internet. DeepNude is a technology advancement that is easier to use and faster to create than deepfakes. DeepNude further rejects the notion that this technology can be utilized for purposes other than claiming control of women’s bodies.
This can enable anyone to remove women’s clothes on the street. Imagine someone taking your photos in the street and being able to undress you via this application.
Advocacy and Reactions
“This is absolutely terrifying,” said Katelyn Bowden, founder and CEO of the revenge porn activist organization Badass. “Anyone can now become a victim of revenge porn without ever having taken a naked photo.” This technology should not be made available to the general population.”
This is an “invasion of sexual privacy,” according to Danielle Citron, a law professor at the University of Maryland Carey School of Law who recently testified to Congress on the deepfake threat.
She explained, “As one deepfake victim told me, it felt like thousands of people saw her naked, and she felt like her body wasn’t her own anymore.”
Software Deployment and Usability
On June 23, DeepNude debuted as a website with a demo of how the software works as well as a downloadable Windows and Linux application. Motherboard downloaded the application and ran it through its paces on a Windows PC. It installed and launched just like any other Windows application and did not require any technical knowledge to use. The output photographs in the free edition of the program are partially obscured by a huge watermark.
The watermark is erased in the premium edition, which costs $50, but a stamp with the word “FAKE” is inserted in the upper-left corner. (It would be extremely simple to crop off the “fake” stamp or remove it with Photoshop.) Motherboard put it through its paces on more than a dozen photographs of women and men in various stages of dress—from totally dressed to string bikinis—and skin tones.
The results vary greatly, but when fed a well-lit, high-resolution photograph of a lady in a bikini squarely facing the camera, the synthetic nude images are passably realistic. The algorithm fills in details such as where clothing used to be, angles of the breasts beneath the apparel, nipples, and shadows with precision.
However, it is not without flaws. Most photos, particularly low-resolution images, produced some visual artifacts. DeepNude completely failed with certain pictures that used unusual perspectives, lighting, or clothes, which appear to throw off the neural network it employs. When we fed it an image of the cartoon character Jessica Rabbit, it deformed and corrupted it completely, transforming stray nipples into a glob of a figure.
Technical Background
DeepNude’s anonymous author, who wished to be identified only as Alberto in an email, told Motherboard that the software is built on pix2pix, an open-source algorithm developed by University of California, Berkeley researchers in 2017.
Pix2pix employs generative adversarial networks (GANs), which function by training an algorithm on a massive dataset of images—in the case of DeepNude, more than 10,000 nude photos of women, according to the programmer—and then competing with itself. This algorithm is comparable to the one used in deepfake films and the one used by self-driving cars to “imagine” road scenarios.
Alberto explained that the algorithm mainly works with women since photographs of naked women are easy to locate online—but he hopes to make a male version as well.
“There are multiple networks because each one has a different task: locate the clothes.” Cover up your clothes. Make up anatomical postures. “Render it,” he instructed. “All of this makes processing slow (30 seconds in a normal computer), but it can be improved and accelerated in the future.”
