Integration of Artificial Intelligence into the US Intelligence System: An Analysis of Application Difficulties

1. Security issue

Artificial intelligence is essentially an extension of computer technology, so whether it is software or hardware, artificial intelligence has considerable vulnerability in the face of malicious behaviors. In view of the uncertainty of the current artificial intelligence technology, the consideration of artificial intelligence security is worthy of attention. In the report of the "Artificial Intelligence Military Application" report, Rand also believes that the machine learning system learned from training data is vulnerable to other types of attacks. Including the so -called data poisoning attack, in this attack, the enemy can manipulate or deceive the training data to affect the expected function of the system. Theoretically, deep learning algorithms need to search historical data that matches the previous malicious code. Therefore, the enemy may deliberately design a targeted algorithm, and only a small part of the code can bypass the test, so as to deceive and seduce our artificial intelligence system to make mistakes or collapse. The "Artificial Intelligence and National Security" report also stated that although humans are not unable to avoid errors, their errors are usually based on individuals, and each error is different. However, the artificial intelligence system may fail in the same way at the same time, which may have large -scale or destructive effects.

In addition, in a highly complex automation system, the system users cannot correctly understand the original intention of the system designer, and then handle the information received improperly, which will also cause faults and cause accidents. On the one hand, the current artificial intelligence algorithm is generally developed for intelligence data or data of a specific type of information. Once the user enters the error data or exceeds the carrying capacity of the system, it is likely to cause failure. On the other hand, for non -professional intelligence personnel, it is almost impossible for debugging and supervising the operation of the artificial intelligence system. Once the autonomous system is captured by malicious programs such as StuxNet, the time for artificial error correction will be compressed. It is difficult for intelligence personnel to stop their self -malicious copying and spread other malicious behaviors in time, which is likely to cause the collapse of the network of their intelligence agency networks or data leakage.

2. Data trap

The American intelligence community has the world's most comprehensive intelligence collection methods, but the data processing capacity is far less than collecting the ability, so there is a large amount of intelligence data processing lag. According to statistics, the U.S. military is equipped with more than 8,000 drones that use good effects. Except for images, it can detect about 1,600 hours of video videos every day. Therefore, the usual views are generally believed that artificial intelligence and big data technology have extensive application prospects in the intelligence community, and the intelligence community will not help constantly seek advanced artificial intelligence and machine learning technologies.

However, in the era of artificial intelligence, the improvement of analysis accuracy has made software have higher requirements for data quality, which challenges artificial intelligence in intelligence work.

On the one hand, well -applied artificial intelligence algorithms in the civilian field are trained by a lot of and standardized data sets, and the data obtained by various intelligence methods are usually incomplete and fragmented. Therefore, the intelligence community can only train the algorithm through extremely limited complete real data. Furthermore, whether the artificial intelligence system can identify and understand these data and it is difficult to judge, and the lack of a certain amount of training samples will make the accuracy of its analysis results worse and unsatisfactory.

On the other hand, the probability of crisis incidents is very low, and each event has its own particularity, which will lead to a huge difference between system warning and real threats. The challenges faced by intelligence work have always excavated high -value signals from a large number of random signals. Although the application of artificial intelligence and big data technology has indeed improved data processing efficiency, the improvement of relying on algorithms is difficult to completely resolve the contradiction between computing power and computing volume.

Therefore, if you blindly promote the integration of artificial intelligence and big data into the intelligence community, it is likely that intelligence agencies will fall into the data trap that blindly pursue the large data volume and ignore the connotation of the data itself. From this perspective, by improving intelligence collection methods and methods and improving the targeted and quality of intelligence collection, it may be an effective way to cooperate with human analysts to improve intelligence work.

3. Technical bottleneck

Artificial intelligence algorithms have liberated intelligence personnel from many complicated transactional work. With the help of big data, cloud computing, and parallel computing technology, artificial intelligence has the computing power that requires a large amount of data when analyzing complex problems. However, with the continuous improvement of data volume, today's computing power can still not meet long -term high load work. In the future, hardware will still be an important factor in restricting the performance of artificial intelligence systems. The development of deep learning technology is closely related to the improvement of GPU, CPU, memory and other hardware performance. There is indeed a mutual promotion role between the system and the hardware. The emergence of parallel computing and the optimization of algorithms can improve the efficiency of the hardware use to a certain extent. Quantum computing based on the concept of quantum theory may be the best solution to the dilemma of computing power. However, this technology is still in the research and experimental stage. It is still unknown whether it can be really put into use.

In addition, artificial intelligence is currently facing another technical issue is human -machine collaboration. To realize the human -machine collaboration of intelligence work, intelligence personnel must understand the artificial intelligence operation mechanism of collaboration, and then further understand its decision -making process and make an evaluation. However, this is more difficult for the neural network that current deep learning algorithms depends on. Using deep learning analysis software is not closer to the human thinking mode than traditional monitoring and early warning methods, it is just because it has stronger computing power and shows high efficiency. In other words, computer analysis paths cannot be close to human thinking about human thinking anyway. Researchers at the Virginia Institute of Technology have created the eye trace system suitable for neural networks, which records the order of computer analysis pixels. Researchers tried to track the decision -making mechanism of artificial intelligence algorithms in this way, but the test exposed the disadvantages of the machine, indicating that even deep learning algorithms could not fully recover the cognitive ability of human analysts.