The counterfeiting cost of AI face changing technology is relatively low and difficult to distinguish. However, can the development of artificial intelligence counterfeit detection technology be a response? At present, AI counterfeit detection technology is still developing, and I hope that in the future, this "cat and mouse game" will become more and more just.
Recently, the news about AI face-swapping scams has gone viral. On April 20, Mr. Guo, the legal representative of a technology company in Fuzhou, suddenly received a WeChat video from a friend, who told Mr. Guo that his friend needed a deposit of 4.3 million yuan to bid abroad.
Because he saw his friend's appearance in the video and heard the voice, Mr. Guo transferred the money with confidence.
But later he found out that he had encountered a "high-end" scam, and the other party used intelligent AI face changing technology to pretend to be a friend to defraud him...
This kind of news really caused a panic in people's hearts.
Because as mortal flesh, it is difficult for you and me to see the flaws in these fake videos with the naked eye.
Doesn't that mean that when encountering such an "AI trick", the chance of ordinary people being deceived is almost 100%?
And slowly, the magic is one foot high, the road is one foot high, since AI face changing has developed to this point, how has the "AI for-fake" technology developed against it?
1. Counterfeiting VS counterfeit detection
When it comes to this problem, the AI counterfeit detection technology team as the "defender" seems to have an unspeakable hardship. Because in this "cat and mouse game" of counterfeiting and counterfeiting, the cost of counterfeiting technologies such as AI face replacement is much lower than the research and development cost of AI counterfeiting technology.
This has led to many enterprises and institutions that develop AI counterfeit detection technology, from the day they decided to fight counterfeiters, they were destined to fight a very difficult war.
1. Face changing technology cost
At present, the cost of AI face changing technology mainly includes the following aspects:
1) The cost of data collection and processing
AI face swapping technology requires a large number of face images or videos as training data, but most of this data can be obtained for free from the Internet, or it can be purchased or customized by payment.
Even sometimes, customizing an AI face changing video is almost "nine pieces of nine free shipping".
2) The cost of algorithm development and optimization
AI face changing technology is mainly based on deep learning and computer vision algorithms, although this process requires certain professional knowledge and skills, but it is not particularly high.
- On the one hand, there are many open source platforms, such as Faceswap, DeepFaceLab, etc., which themselves provide ready-made algorithm code, and users only need to download and use it, without self-development or improvement.
- On the other hand, even if users need to develop or improve their own algorithms, they can learn from existing algorithm frameworks and ideas without starting from scratch.
3) The cost of hardware equipment and operating environment
In general, a slightly better equipped personal computer can meet the needs and does not require particularly expensive professional equipment.
For example, DeepFaceLab, a popular AI face-changing software, requires only one NVIDIA RTX series or GTX series discrete graphics card with 6GB of video memory to run.
The cost of such a personal computer is about thousands of dollars.
Compared with the various costs required by AI counterfeit detection technology, the cost of these counterfeiting technologies is simply as low as child's play.
2.AI The cost of counterfeit detection
Specifically, the cost of AI counterfeit detection roughly includes the following parts:
1) The cost of data labeling and screening
AI counterfeit detection technology requires a large number of images or videos as training data, which need to be manually labeled and screened.
However, the difficulty of this process may be beyond the imagination of many people.
For example, if you want to annotate an hour's video for face recognition, you need to first divide the video into frames, and then locate, identify, classify and other operations on the faces in each frame, and also consider the occlusion, blurring, lighting and other influencing factors of the face, as well as the face's posture, expression, age and other attribute information.
As a result, an hour's worth of video can take dozens or even hundreds of hours to finish annotating.
Secondly, data annotation is not only simply labeling images or videos, different tasks and scenarios require different labeling rules and methods, and annotators need to master relevant domain knowledge and technical capabilities to ensure the quality and validity of data.
2) The cost of algorithm design and testing
AI counterfeit detection technology is mainly based on deep learning and computer vision algorithms, which need to be designed and tested according to different types and quality of counterfeit content to improve the coverage and robustness of detection.
Algorithm design and testing is a complex and continuously updated process that requires a dedicated team and resources. The cost of algorithm design and testing depends on factors such as the complexity, performance, and efficiency of the algorithm, but in general, the research and development cost of each algorithm project is at least between hundreds of thousands and millions of dollars.
3) The cost of hardware equipment and systems
But anyone who has understood the cost of large model training knows the high hardware requirements of this process. Even if today's machine vision technology has been optimized layer by layer, the price of a complete machine vision system is at least tens of thousands to hundreds of thousands of yuan, including hardware costs, software costs, computing costs and storage costs.
To sum up, due to the natural cost advantage of AI counterfeiting technology, the current war of counterfeiting VS counterfeiting presents a situation of "regular army encircling and suppressing guerrillas".
In such a contest, the "disaster areas" that suffer the most are often places that are difficult to cover with high-cost AI counterfeit detection technology.
Second, the weakness of counterfeit detection technology
Looking back at the AI fraud incidents that have been seen in major media over the years, it is not difficult to find that the most common victims of such incidents are often ordinary people as individuals, rather than enterprises, governments or financial institutions.
For example, on May 22, Mr. He from Anqing, Anhui Province, received a video call from an acquaintance asking him to help transfer an account. "Because I made a video call and I was an acquaintance, I didn't think much about it, so I transferred the money." At the time of the incident, Mr. He relaxed his vigilance and transferred 2.45 million yuan to the other party.
Similarly, in March 2023, a woman in Shaanxi was scammed by scammers using AI face-changing technology to impersonate her daughter and defrauded her of a lot of money through a video call. The scammer first added the woman as a friend through WeChat and claimed to be his daughter's classmate.
After that, the scammer replaced his face with the face of the woman's daughter through AI face changing technology, and defrauded another sum of money on the grounds that her daughter had a car accident.
In fact, it is not difficult to understand the reason for this phenomenon, because the current research and development cost of AI counterfeit detection technology is relatively high, so most companies with such technologies are only open to corporate, government or financial customers, and cannot cover ordinary individual users.
Taking Baidu as an example, although it has launched a face depth authentication interface on its own AI open platform to accurately identify the face images generated by Al face changing technology, there is a certain threshold for the use of this interface.
Users need to register an account on the Baidu AI Open Platform and apply to use the interface. Secondly, the interface is chargeable, and there are different charging standards according to the number of calls and the quality of service.
Similarly, Baidu's product Du Xiaoman, although it has also launched its own anti-deep counterfeiting technology, but this technology mainly covers the group of customers in the financial field, and the scenarios involved are mainly lending, investment, wealth management and other businesses.
Although major enterprises, based on micro-expression analysis, graph convolution (GCN) technology and self-supervised pre-training methods based on reconstruction, the current AI counterfeit detection technology has been greatly upgraded. However, as large models become more popular, AI fraud technology has also begun to show a trend of cheapness and universality.
Unless these rigorous and efficient counterfeit detection technologies can be popularized on every individual user's WeChat, and call them like mini programs without threshold and difficulty. Otherwise, the kind of AI technology that scams through "WeChat friends" will always have an opportunity.
Things have developed to this point, and the entanglement between AI technology and ethics and law seems to have reached a more critical turning point.
If these low-cost, AI fraud incidents targeting individuals are repeatedly banned, and even take advantage of this wave of large model "east wind" to further sweep the country. Then there will be more lawbreakers who are unscrupulous and concoct the law, which will cause further panic among the people.
In the case that AI counterfeit detection technology cannot be sunk, in order to prevent the frequent occurrence of such incidents, the regulatory authorities can only further tighten the review mechanism for the use of AI technology, and the strict review mechanism will indirectly reduce the activity of the AI ecosystem and the prosperity of the market.
Many ideas or projects that could have worked wonders may not even have begun, but they are stuck in the complicated review stage.
But in addition to this pessimistic outlook, there is another possibility that more companies and institutions, under the pressure of such AI scams, are beginning to seek breakthroughs and innovations, forming an innovation-driven mechanism.
Third, the demon is one foot high, and the road is one foot high
Although AI counterfeit detection technology is facing a situation of high cost and difficulty in sinking, the way to break the game is not non-existent. Because the upgrade of a technology, the impact may not be vertical and single, but often may be multidimensional and vast.
For example, the underlying technologies such as computer vision and image recognition involved in AI counterfeiting have a very broad application range and huge development space.
For example, in terms of intelligent driving, the current computer vision, image recognition and other technologies can not completely replace human driving, but also need to combine other sensors, radar, maps and other information to improve the accuracy and robustness of perception, but also need to solve the complex and changeable situation of the road, improve the intelligence and safety of decision-making.
In the field of smart agriculture, by improving the coverage and timeliness of image recognition, more comprehensive and real-time monitoring and diagnosis of crops, soils, diseases and pests can be achieved, thereby improving the quality and efficiency of agricultural production.
Similar areas of benefit include transportation, medical care, entertainment and so on.
In the longer term, the technological spillover caused by this sinking process will enable more economic actors to acquire and use advanced technologies, thereby promoting the diversification and vitality of innovation activities.
At the same time, the reduction of technical threshold will inevitably lead to a decline in production costs and prices, enabling enterprises to further expand the market scale and promote economic growth.
From this point of view, the current AI fraud incident is only a kind of disaster or grinding that the large model must suffer in the process of sinking. Only after such a test can the various technologies brought by AI be truly widely accepted by the public in law and ethics.
Author: Famous Ear
Source public number: AI new intelligence
This article is published by the media that everyone is a product manager cooperation media @AI New Intelligence, and it is forbidden to reprint it without permission.
The title image is from Unsplash and is based on the CC0 protocol
The views of this article only represent the author himself, everyone is a product manager, the platform only provides information storage space services.