Is AI fraud a 100% success rate? Anti-deepfake model 'defeat magic with magic”

A few days ago, a topic about AI fraud with a success rate of close to 100%# became a hot search topic on Weibo. The legal representative of a technology company in Fujian was defrauded of 4.3 million yuan in 10 minutes because he watched a video made using AI face-changing technology.

An AI-related scam also occurred abroad. An email with a video of Google CEO attached caused many YouTube bloggers to download files containing dangerous viruses.

Both of these fraud incidents involved deepfake technology. This is a face-changing method that has been around for 6 years. Nowadays, the explosion of AIGC technology has made it easier and easier to create hard-to-identify deepfake videos. For the financial industry where facial recognition is widely used, preventing deep fake attacks is also particularly important.

In the financial industry, the fraud caused by deepfake is mainly identity fraud, that is, using deep fake images and videos to impersonate other people's identities, deceiving the identity verification system in the financial credit process, and then committing fraud and malicious registration. . At present, the financial industry has relatively mature technical methods and solutions for dealing with deepfake, and Du Xiaoman has accumulated rich experience in dealing with deepfake.

Du Xiaoman introduced that in recent years, the trend of using deep fake technology to bypass the face recognition process has increased, posing a certain threat to the real-name authentication system of financial institutions. Developing an “anti-Deepfake” detection algorithm and handing over content authenticity verification to artificial intelligence is the most reliable method. The Du Xiaoman detection model algorithm strategy to prevent deep forgery successfully solves the problem of creating fake videos from three aspects.

The first is to generate defects. Specifically, due to the lack of relevant training data, the deepfake model may not be able to correctly render some human facial features, ranging from abnormal blinking frequency to inconsistency between mouth shape and voice, etc. By designing specific analysis algorithms, the detection model can extract those "basically visible to the naked eye" features and perform analysis and judgment.

The second is the inherent attribute. Since different cameras have different device fingerprints, models like GAN will also leave unique fingerprints for identifying the generator when generating faces, so clues can be found through comparison.

The third detail is high-level semantics. It refers to issues such as detecting the coordination of facial action units (muscle groups), the orientation consistency of various facial areas, and the microscopic continuity of videos. Because these details are difficult to model and copy, it is easy to get caught.

Of course, since a single feature is difficult to adapt to complex deepfake content, the overall framework of the detection model adopts multi-feature fusion to ensure the robustness of decision-making.

In addition to the advantages of data samples, Du Xiaoman also incorporates his own originality, including neural network search and optimization algorithms, micro-expression analysis and graph convolution (GCN) technology, and reconstruction-based self-supervised pre-training methods, allowing The model realizes the transformation from "counterfeiting" to "authenticity".

It is precisely because of this that Xiaoman’s anti-deep fake detection model successfully passed the special face recognition security evaluation of the Academy of Information and Communications Technology in September last year, and obtained the excellent certification for live detection security protection capabilities. In terms of specific effects, it can cover various forms of deep fakes, including static portrait picture activation, AI face changing, false face synthesis, etc., achieving a recall of more than 90% with a false alarm rate of one thousandth, which is an accuracy of 99% .

As new deepfake tools continue to emerge, the financial industry will face an increase in deepfake attacks. Du Xiaoman believes that more counterfeiting detection technologies in the future should focus on mining semantic features, cross-modal features, etc., so that the model can use high-level semantics with strong interpretability to detect counterfeiting.

(Source: Guangming.com)

The above is the detailed content of Is AI fraud a 100% success rate? Anti-deepfake model 'defeat magic with magic”. For more information, please follow other related articles on the PHP Chinese website!