- Статьи
- Science and technology
- Checking for light: how optical computing can change the fight against deepfakes
Checking for light: how optical computing can change the fight against deepfakes
The more convincingly artificial intelligence fakes reality, the harder it is to trust what it sees. Scientists at the University of California at Los Angeles have proposed searching for traces of digital deception using light. Whether this technology will be able to protect users from deepfakes and how it will change the future of computing is in the Izvestia article.
How light helps to detect deepfakes
The UCLA researchers' development combines a digital neural network and an optical system. At the first stage, the program analyzes video frames and converts the selected features into a special phase distribution, which is displayed on a spatial light modulator. Then, due to the propagation of light, a signal is generated that makes it possible to classify the recording as real or fake.
In an experiment with 15 simultaneously processed videos, the average accuracy was 97.79%, sensitivity was 99.86%, and specificity was 95.72%. The last two indicators reflect the ability to detect forgeries and correctly recognize real records, respectively. The authors consider the system as a preliminary filter, after which suspicious content can be sent for more detailed verification. The study is published in the journal eLight.
The main technological feature of the work is the parallel processing of independent video streams, explained Mikhail Petrov, a leading researcher and deputy director of the ITMO Institute of Photonics. At the same time, the computationally expensive part of the analysis remains for the digital neural network.
"First, the digital neural network analyzes video frames and encodes the features into a phase distribution convenient for optical processing, and then the light performs decoding due to diffraction and interference and generates the classification result," he said.
Thus, optics helps to perform the final stage of recognition. The final result still depends on how well the digital part has identified the signs.
The difficulty lies in the fact that such signs change along with generation technologies. According to Oleg Rogov, head of the scientific group "Trusted and Secure Intelligent Systems" at the AIRI Institute, the detector often learns to find traces of a specific generator: errors on the border of the superimposed face, mismatch of lighting, characteristic noise or violations of facial expressions between frames.
"The new generation architecture leaves other traces or almost none, and the model, accustomed to the old features, begins to make mistakes," the expert explained.
Fully synthetic videos are especially difficult to detect, where the original recording with the replaced face is missing. In addition, compression on platforms can erase high-frequency features that are useful for recognition. Therefore, the development of equipment by itself does not solve the problem, even a high-performance detector needs to be adapted to changing content.
What does development change for computing technologies?
The practical interest in optical architecture is primarily related to the ability to distribute processing costs across multiple threads. For services that check large amounts of video, the cost of analyzing each recording is important. If it can be reduced without significant loss of quality, the test can be applied more widely.
According to Mikhail Petrov, in the proposed architecture, the power consumption of the optical decoder is about 86-95% lower than that of the digital analog chosen by the authors. However, for the entire system, the savings amount to about 13%: most of the energy is spent on digital processing.
"This does not automatically mean that optics have an advantage over any fully digital neural network specifically optimized for the same task," the specialist emphasized.
The comparison refers to specific architecture options. To assess the prospects for implementation, it will be necessary to take into account the entire process: personnel training, modulator operation, lighting and result reading. Similarly, it is necessary to compare the speed and cost of the equipment.
According to Petrov, scaling capabilities depend on the size, resolution, and performance of spatial light modulators. An additional limitation is crosstalk between channels: increasing the number of simultaneously checked records requires maintaining sufficient independence of their processing.
The transition to an industrial device also involves solving the problems of compactness, stability and integration of optics. But even a technically sound system will require an application in which the benefits justify the development and maintenance costs.
Hardware upgrades are a separate issue. The basic scheme can be adapted by retraining the digital part, if this is enough to maintain accuracy. However, in variants with additional static diffraction layers, a significant change in the task may require the replacement of physical elements, the expert explained. The use of additional programmable modulators will make the system more flexible, but will increase the cost.
In addition, the prospects of the direction go beyond the detection of fakes. According to Petrov, optical accelerators are being developed for vector-matrix operations, on which neural networks are based, as well as for optimization and combinatorial tasks. Research is also underway in Russia, including at ITMO.
Where detectors can improve safety
For businesses and users, the importance of new technologies is determined by whether they can help detect fraud before transferring money or confidential information.
According to Dmitry Anikin, head of data research at Kaspersky Lab, video chips are already being used in fraudulent scenarios involving the substitution of company executives. For ordinary users, threats include misinformation, false financial recommendations, and fake video messages in messengers.
Automatic recognition can become an additional level of protection. However, its effectiveness depends on what actions the organization or user takes after receiving the warning.
"Automatic verification can help combat the threats of fake videos, but the processes in which this will be implemented are primary," Anikin said.
For companies, this means the need for protocols that make illegal transfers more difficult. Even a convincing video message should not cancel the stipulated approval procedure. The result of the detector in such a system may become the basis for additional verification.
The cost of an error varies depending on the application, the expert added. In a corporate financial process, additional confirmation may be justified. In a mass user service, constant erroneous warnings can lead to complaints and a deterioration in the quality of the application.
Separate requirements arise when analyzing video calls. As Anikin explained, basic recognition technologies may be similar to checking ready-made recordings, but embedding a model in a user conversation raises the question of the sensitivity of the processed data.
Oleg Rogov also draws attention to the need to test the system in terms of a specific platform. In the real stream, there are re-compressed videos, screen recordings, shooting a monitor with a phone, vertical videos with filters and superimposed text. Resistance to certain types of noise and blur does not allow us to assess all these situations in advance.
Tests against targeted attempts to circumvent protection are also necessary. According to Rogov, the physical implementation of the optical part may complicate the attack, but the digital encoder remains a potentially vulnerable component.
Therefore, the implementation should be preceded by a long-term pilot on real traffic. It will allow you to estimate the frequency of false positives and monitor how the quality changes as new types of content become available.
Why accuracy does not guarantee trust
The main limitation of laboratory results is manifested during the transition to mass use. The 95.72% specificity means that in the conditions of the corresponding test, about 4.28% of the real records were mistakenly classified as fake.
"About every 23rd real video is marked as fake, and on a platform with millions of downloads per day, this is a huge stream of false positives," Oleg Rogov noted.
This result does not determine the error rate on any platform, but it shows why general accuracy alone is not enough to make an implementation decision. According to the expert, data on different types of manipulations and compression ratios, confidence intervals, as well as tests on independent sets of videos are needed.
In the main experiment, the researchers used 210 unique test videos and repeatedly changed the sample of frames. This allows you to check the stability of the result to the composition of the input data, but does not increase the diversity of the records themselves. Rogov also considers it necessary to assess whether the model remembers the features of faces and shooting conditions instead of signs of forgery.
The expert interprets the recognition of Veo 3 videos after further training as evidence of the ability to adapt to a certain generator. This result does not prove the willingness to discover fundamentally new ways of creating videos.
—A realistic goal is to reduce the cycle of retraining to days and build an architecture that honestly communicates its insecurity on unfamiliar data," Rogov believes.
According to him, mass automatic primary filtering is realistic in the next two to three years. However, final decisions will require a combination of several tools: verifying the origin of recordings, labeling synthetic content, detectors, and human involvement in controversial cases.
At the same time, information about the origin helps to establish the history of the file, but by themselves does not confirm the authenticity of the captured event. The absence of a label also cannot be automatically considered proof of forgery.
The effectiveness of such a system depends on the coordinated operation of generators and platforms, Rogov emphasized. Optical computing can make one of the verification stages more economical, but the tasks of updating models, error handling, and source confirmation will remain. It is the results of long-term tests on real content that will show what place the new architecture can occupy in this infrastructure.
Переведено сервисом «Яндекс Переводчик»
