Home GADGETS Facebook’s latest AI can not only detect deep fakes, but also know where they come from

Facebook’s latest AI can not only detect deep fakes, but also know where they come from



even Deep Tom Cruise It is believed that the latest generation of digitally manufactured faces has stepped out of the uncanny valley and has been approaching photo realism. While the possibilities for entertainment using this tech are boundless, deep fake videos have the potential to severely disrupt the public’s trust in government and our elected officials — even the ability to believe our own eyes. On Wednesday, Facebook and Michigan State University launched a The novel method can not only detect depth forgery, but also discover which generative model produced it by reverse engineering the image itself.

In addition to telling you whether an image is a deep forgery, many current detection systems can also determine whether the image was generated in a model that the system saw during training-called “near set” classification. The problem is that if the image is created by a generative model that the detector system has not trained, then the system will not have previous experience to identify fakes.

The team explained in a blog post on Wednesday that the FB-MSU reverse engineering technology, although not exactly a cutting-edge method, “relies on revealing the unique patterns behind the AI ​​model used to generate a single deep vacation image.”

“We started with the image attribution, and then worked on discovering the attributes of the model used to generate the image,” the team continued. “By generalizing image attribution to open set recognition, we can infer more information about the generative model used to create deep forgeries, rather than just recognizing things that have never been seen before.”

More importantly, the system can compare and track the similarity of a series of deep forgeries, allowing researchers to trace the forged image groups back to a single generation source, which should help social media moderators to better track coordination errors Information activities.


In order to implement this detection technique, FB-MSU researchers first ran a set of deep vacation images through a fingerprint estimation network. FEN is able to distinguish the subtle patterns printed on the image by the specific equipment that made the image. For digital photos, each of these patterns is unique due to differences in camera manufacturing. The same is true for deep fraud-each generative model has its own quirks, which are imprinted in their creations and can be used to reveal the identity of the model based on the image itself.

Since there are actually an unlimited number of generative models in the Internet field, researchers have to generalize their search for fingerprints of these images. The team explained: “We usually use different constraints to estimate fingerprints based on fingerprint attributes, including fingerprint amplitude, repeatability, frequency range, and symmetrical frequency response.” These constraints are then fed back to FEN, “to force the generated fingerprint to have These required attributes.”

Once the system can consistently distinguish real fingerprints from deeply forged fingerprints, it will dump all these fake fingerprints into the analytical model to infer their various hyperparameters. The hyperparameters of the generative model are the variables it uses to guide its self-learning process. So, if you can figure out what the various hyperparameters are, you can figure out what model uses them to create the image. The Facebook team likened it to the ability to recognize the various engine components of a car simply by listening to idle sounds.

Since the FB-MSU team is involved in unknown research areas in this study, there is no specific baseline to compare their test results. Therefore, the team created its own and found that “compared to random vectors with the same length and distribution, there is a stronger and broader space between the generated image and the embedding space of meaningful architectural hyperparameters and loss function types. Relevance.” So basically, they can’t say objectively how good their system is, because there is actually no other research to compare with it, but they do know that it is more effective than blind luck.

All products recommended by Engadget are selected by our editorial team, independent of our parent company. Some of our stories include affiliate links. If you purchase goods through one of these links, we may earn member commissions.


Source link


Please enter your comment!
Please enter your name here