With the rise of artificial intelligence technology, the research in this field has been deepened, and its application has already been integrated into many fields and has become an inseparable part of it. At present, artificial intelligence has covered the knowledge of multiple subjects such as mathematics and computer science, and there is a detailed technical discussion process for each direction. There has been a single spark and a prairie fire. But artificial intelligence technology is a “double-edged sword”, and when it brings convenience to our lives, it will also bring related problems. At present, security issues have received the most attention. Starting from its basic concepts and connotations, this article mainly combines some cases to analyze the current security risks of artificial intelligence technology and make relevant recommendations at the end.
Basic concepts of AI
In 1956, at the Dartmouth Conference in the United States, scientist John McCarthy first proposed the term “Artificial Intelligence” (AI). Artificial intelligence is a technology that makes the behavior of machines look like intelligent behaviors displayed by humans. When the concept of artificial intelligence was proposed, scientists mainly determined the criteria and research goals of intelligence, but did not answer the specific connotation of intelligence.
Since then, many well-known scholars have put forward their own opinions on the connotation of artificial intelligence, reflecting the basic ideas and basic content of artificial intelligence: research on the basic theories, methods and technologies of how to use computers to simulate human intelligent behavior. However, due to the continuous evolution of the concept of artificial intelligence, a unified definition has not yet been formed.
On the one hand, the connotation of artificial intelligence security includes the application of artificial intelligence in the security field, such as artificial intelligence assisting or replacing humans in more and more industries. On the other hand, the security problem of artificial intelligence itself is the focus of this article. How to promote the application of artificial intelligence to be safer and more ethical has always been a long-term thinking and deepening proposition for human beings. The immaturity of artificial intelligence technology at this stage will lead to many security risks. For example: When an autonomous car has to make a choice between hitting an old man and hitting a child, how will it choose? What is the reason behind it? Which factors will it pay more attention to when making choices? Can it justify itself? This is just the result of mathematical operations, which is inexplicable in the eyes of humans. In addition, the strong dependence of data and the malicious application of man-made are security risks caused by the immaturity of technology; furthermore, not only artificial intelligence systems need security, but other systems also need security. Using artificial intelligence to attack other systems and detect vulnerabilities in other systems, this behavior may bring security risks to cyberspace and national society .
Therefore, the connotation of artificial intelligence security includes : reduce the security risks caused by the immaturity of artificial intelligence technology, and at the same time use artificial intelligence to build a security protection system to reduce the harm caused by man-made malicious applications.
Before the analysis, we first briefly introduce the application of artificial intelligence. Because artificial intelligence has outstanding advantages in data analysis, knowledge extraction, and autonomous learning, it is widely used in network protection, data management, information review, intelligent security, financial risk control, and public opinion monitoring. In these aspects, there are often several security risks. Common risks are as follows:
In recent years, the well-known deep learning frameworks TensorFlow, Caffe, etc., and their dependent libraries have been found to have security vulnerabilities many times. These vulnerabilities can be exploited by attackers and cause system security problems.
Take Generative Models as an example . The original working principle is: input x is mapped to a low-dimensional representation through an encoder, and then mapped back to a high-dimensional representation of z through a decoder, which is represented by a representation.
In addition, artificial intelligence may be used to write computer viruses and Trojan horses. The original malicious scripts are manually written by humans. Artificial intelligence technology can automate these processes and bypass security detection by inserting adversarial samples . Similarly, artificial intelligence technology can also automatically generate intelligent botnets . These networks can attack other systems on a large scale and automatically without waiting for the command of the botnet controller, which greatly aggravates the network attack.
Attackers can reversely obtain network training data sets through the internal parameters of the network ; artificial intelligence technology will also enhance the ability of data mining, which increases the risk of privacy leakage, such as the Facebook data leakage incident in March 2018.
Inaccurate, unreasonable or incorrect definition of the objective function of the deep learning network may also lead to wrong or even harmful results. The wrong objective function, the objective function with too high computational cost, and the network with limited expressive ability may cause the network to produce wrong results. For example, in March 2018, Uber’s self-driving car failed to recognize pedestrians on the road in a timely manner, resulting in collisions with pedestrians and pedestrian deaths.
The prejudice of algorithms and the inexplicability of artificial intelligence are also a major problem. In the past, the United States used artificial intelligence algorithms to predict criminals. The list was exposed. Many of the innocent people were wronged, and most of the wronged people were blacks, and the system in the decision-making algorithm made such a decision, even if it was developed. The author cannot give a reasonable explanation.
The appearance of adversarial samples will also make the algorithm misjudgment. Add a small amount of noise in the following figure, and AI will recognize the panda with a high degree of confidence in the gibbons.
As long as there is enough training data, artificial intelligence can produce false information for illegal activities. For example, AI face-changing technology DeepFakes, and DeepNude that appeared some time ago. Some criminals use fake voice and video to commit fraud. At present, the chat robot invented by Google has completely deceived human beings when talking with people on the phone.
This is also the hottest and most controversial aspect. First of all, artificial intelligence will replace workers in many positions in the near future. This is correct from the current development trend. If people in relevant positions do not learn new skills, they will face unemployment, which will lead to serious social problems. Secondly, artificial intelligence systems may endanger human safety. Due to the immaturity of technology, some behaviors such as autonomous vehicles and drones may lead to human safety risks. In 2016, Tesla was unable to recognize a white truck on a blue sky background and caused a car accident; in 2018, due to the failure of the machine vision system to recognize pedestrians that suddenly appeared on the road, the Uber automatic car collided with pedestrians and caused pedestrian deaths; in recent years, there have been more accidents in China. A drone interference led to an emergency landing of the flight. In addition, artificial intelligence products may have an impact on existing social ethics :
1) The bias of AI algorithms will affect social fairness and justice. For example, why is the crime rate of blacks so much higher than that of whites?
2) The AI algorithm lacks ethical constraints, and only knows to achieve its own goals without considering the rights and interests of others. For example, some apps will use artificial intelligence systems to automatically send text messages to users to promote products and advertise based on some recent activities of users. It cannot guarantee the authenticity of its content and whether its actions violate certain rights of users.
3) If humans and artificial intelligence have emotions like husbands and wives, fathers and daughters, modern ethical norms will be examined.
Artificial intelligence can be used to influence public ideology. For example, in the US general election, peer-to-peer fake news is used to push candidates to influence public ideology and affect the results of the general election. Artificial intelligence can also be used to build new military forces. For example, in August 2018, the President of Venezuela was bombed by drone bombs in a public event. This is the first case of terrorist activities using artificial intelligence products in the world.
As a strategic technology leading the future, artificial intelligence has become an important engine driving the development of various fields of society. However, AI is a “double-edged sword,” the need to strengthen the security risks and prevention research, and promote their application in the security field, trying to wave the development of artificial intelligence in the next round of the accounts have opportunities to win the initiative. For this reason, there are several suggestions for the development of artificial intelligence security: