Currently, the creation of strong artificial intelligence (eng. Strong Artificial Intelligence (AI)) to ensure the required cybersecurity of digital platforms Industry 4.0 is one of the most interesting scientific and technical problems of our time. In the 1940s, when Norbert Wiener's book Cybernetics, or Control and Communication in the Animal and in the Machine, and other scientific papers on this topic were published, when the first computers of the von Neumann architecture appeared and began to be distributed. The mentioned problem was transferred from the field of science fiction to the field of real theoretical research and engineering developments. Since then, experts in the field of cyber security have been eagerly awaiting the emergence of fundamentally new technical information protection systems, the level of intelligence of which will be comparable to that of humans. That is, such engineering solutions, the distinctive ability of which will be the independent association and synthesis of new knowledge. Let's take a brief look at the history of the issue and dwell in more detail on the possible formulation of tasks for creating strong cybersecurity artificial intelligence.
In 1960, at the Massachusetts Institute of Technology under the guidance of John McCarthy, the
first functional Lisp programming language was created, based on the theoretical foundation of the
lambda calculus by the famous mathematician Alonzo Church (1903–1995) [
In the early 1980s, Japanese professionals started developing a so-called fifth-generation computer
with advanced AI functions. By that time, Japan had achieved a significant success in the automotive
and aviation industries, and intended to reach a new level of technological development. In the fact they
were supposed to develop a new architecture of parallel computing systems (Figure 2) with a
recordsetting performance of 100 million -1 billion LIPS. At that time, the computer performance was about
100 thousand LIPS, where LIPS is a logical inference per second [
The features of the Fifth generation computer are listed below: • New computing system architecture (not von Neumann); • New microcircuit production technology, which marks the transition from the silicon to gallium arsenide, increasing the speed of the main logic elements; • New methods of information input-output - recognition and synthesis of speech and images; • Rejection of traditional algorithmic programming languages (Fortran, Algol, etc.) in favor of functional Lisp and logical Prolog programming languages; • Focus on the tasks of AI with automatic search for solutions based on logical inference.
The corresponding State program was launched in order to achieve the goals in Japan (1982-1992)
[
program for creating MARS and Kronos processor supercomputers (1985-1988).
In the mid-1980s, the expert systems became widespread (see the excellent book by Eduard
Viktorovich Popov), which were intended to replace the specialists in various subject areas. The
classical expert system was a program based on the “if - that” (the rules of the Post), and allowed to
recognize the situations and draw the simple logical conclusions. Hundreds of such expert systems were
developed, including Expert, Expert-PRO, GURU, etc [
In the 1990s, the relatively new models and methods of neural networks and genetic programming replaced the logical programming.
As a rule, an Artificial Neural Network (ANN) (Figure 3) is understood as a mathematical model, as well as its software and hardware implementation, based on the principles of organization and functioning of the biological neural networks - nerve cells of a living organism [Ошибка! Закладка не определена.6-32,37-44]. For example, the modifications of the first neural networks of W. McCulloc and W. Pitts, who have found an application in the pattern recognition problems, in control, prediction, imparting properties of adaptability and self-organization, and etc.
From an engineering point of view, an ANN is a system of the relatively simple processors (artificial neurons) that receive and send signals to each other. At the same time, the neural networks are not programmed in the usual sense of the word, but are learnt. Here the opportunity to be learnt is one of the main advantages of neural networks over traditional algorithmic systems. Technically, learning is to find the coefficients of connections between neurons. In the process of learning, the neural network is able to detect the complex dependencies between the input and the output data, as well as perform a generalization. This means that in case of successful learning, the network will be able to return the correct result, based on data that was missing in the learning sample, as well as in the partially distorted data (incomplete and/or “noisy”) (Figure 4).
Let us note that the basic models of the neural networks have been known since the late 1950s, but
they became widespread after the development of the backpropagation, which allowed training the
multi-layer neural networks. Such multilayer networks in which there was at least one intermediate
(“hidden”) layer of neurons between the input and output layers can be trained how to perform a much
larger number of functions, compared to their simpler predecessors. In combination with the computer
technology achievements and the supercomputers’ construction, this allowed the construction of the
first neural networks, which quite successfully solved, among other things, the cybersecurity problems.
(Figure 5 and Figure 6) [
Genetic programming is a type of evolutionary computing method. Here, some initial populations (data structures and/or data processing programs) are considered as initial data. As a result of the random mutation and reproduction (“crossing”), the new populations appear. At the same time, a certain selection criterion (fitness function) allows selecting the best solutions. As Nick Bostrom had correctly noted, “In practice, however, getting evolutionary methods to work well requires skill and ingenuity, particularly in devising a good representational format. Without an efficient way to encode the candidate solutions (a genetic language that matches latent structure in the target domain), evolutionary search tends to meander endlessly in a vast search space or get stuck at a local optimum.” At the same time, the evolutionary computations require the significant computational resources.
In practice, in order to apply the models and AI methods in cybersecurity (Figure 7), the special software tools may be needed. These include the open source libraries, ready-made applications, such as the Gigster platform, as well as Microsoft Azure Machine Learning cloud services, Amazon Machine
Microsoft have opened the third-party developers an access to their AI-bots to integrate the voice commands into applications. Also, there are a number of functional platforms, such as Datanomiq, a data science startup based on SAP solutions and services, as well as a number of open source AI application libraries, including the Microsoft Cognitive Toolkit. (Figure 8).
Also, in order to build a multilayer deep neural network, you can apply the capabilities of the DGX1 supercomputer from NVIDIA, which allows more than 12 times increase the performance of learning tasks, compared to the classical architecture of the "von Neumann" computer. At the same time, the library of the DGX-1 programs3 will significantly simplify the process of developing the Deep Learning applications. Let us note that the library includes the NVIDIA Deep Learning GPU Training System (DIGITS)4, a full-featured interactive system for creating the Deep neural networks (DNN), and a GPU3 https://developer.nvidia.com/deep-learning#source=pr 4 https://developer.nvidia.com/digits#source=pr accelerated library of primitives for creating DNN - the NVIDIA CUDA Deep Neural Network (cuDNN). In addition, the system contains a number of optimized frameworks for deep learning - Caffe, Theano and Torch. DGX-1, etc.
In the 2000s, in the developed countries (USA, EU countries, China, Russia and others) a new technological structure of society was formed on the basis of so-called convergent NBIC technologies. For example, in the United States, a program of the National Science Foundation and the Department of Commerce under the NBIC - Nanotechnology, Biotechnology, Information technology and Cognitive science is being implemented. In the European Union, the following programs are being implemented:
Neurons, Genes). China has launched a similar China Brain program. The national technology initiative Neuronet had started development in Russia (CoBrain or Web 4.0 program). Under this program, a number of leading national research and production companies, research institutes and universities, including OJSC Radar system Technology Information (RTI), Research and Development Center of Kurchatov Institute, Research Institute of Neurocybernetics named after A. Kogan, Military Space Academy named after AF Mozhaisky, Moscow Institute of Physics and Technology, St. Petersburg Electrotechnical University “LETI”, National Research University of Information Technologies, Mechanics and Optics, started the pilot production of hybrid and artificial biosimilar materials, technical systems of bionic type and technological platforms based on them. In the future, it is planned to create the complex anthropomorphic technical systems and “nature-like” technologies, combining the components of animate and inanimate nature.
The term cognitive comes from the Latin word cognitio (cognition). The improvement of mathematical models of thinking processes contributed to the development of a cognitive approach in the technical field. The first “artificial cognitive systems” appeared, representing “intelligent” software and hardware systems based on the traditional architecture of the Hungarian-American mathematician and physicist John von Neumann.
The prerequisites of the modern cognitive approach were the fundamental results [
The core of the modern cognitive approach is the methods of cognition, perception and information accumulation, as well as methods of thinking or using this information for the “judicious” solution of the problems. It is believed that artificial cognitive systems are able to "repeat" the complex behavioral functions of the nervous system and even the work of the human mind.
Modern studies of the cognitive systems are conducted on the basis of the neurophysiological principles of the nervous system construction and the cognitive methods of human cognitive and mental activity. For example, in the work of L. A. Stankevich “Artificial cognitive systems” the use of artificial cognitive systems with hybrid architectures in robotics is justified. At the same time, a cognitive system is defined as a system that is capable of learning about its environment and adapting/changing it, due to the accumulated knowledge and acquired skills in the operation process. Two main types of artificial cognitive systems are clearly distinguished: the cognitive and emergent ones.
The actual cognitive systems include: • Traditional character systems (Allen Newell and Herbert Simon); • Systems, based on the theory of cognition, which applies training and the acquisition of symbolic knowledge (J. Anderson);
• Systems, based on the theory of practical reason and high-level psychological concepts of persuasion, plans and intention (Michael Bratman).
Here the former are capable of generating some character structures or expressions. In this case, a symbol is a physical pattern that represents a certain component of an expression (or a character structure). The second ones are based on a system of products and a generalized model of human thinking and knowledge, containing memory, knowledge, decision making, and learning. In this case, the learning contains declarative and procedural steps, depending on the student knowledge. Others implement a decision-making process similar to the traditional practical conclusion.
The emergent systems consist of: • Connectionist systems; • Dynamic systems; • Inactive systems.
The former implements the parallel processing of the distributed activation patterns, applying the statistical properties, rather than logical rules. The latter study the various self-organizing motor systems and human perception systems, examining the relevant metastable behavioral patterns. For others, the definition of a cognitive entity, that is, a purposeful behavior of the system, occurs when they interact with the environment.
Thus, the general methodology for the development of hybrid cognitive technical systems was proposed and substantiated:
• Formalized cognitive concepts and methods for creating the effective self-learning and selfmodifying systems; • Methods for the synthesis of the original cognitive components (modules and networks of modules) capable of accumulating knowledge through training and self-learning. At the same time, the components are built on the basis of a combination of neurological, immunological and triangulation adaptive elements that are most effective for multidimensional functional approximation, as well as corresponding behavioral networks;
• Methods for implementing the cognitive components and systems, based on specially developed software. The software implementation of cognitive components is based on the original models of information processing and training, and cognitive systems are based on multi-agent technology. This cognitive multi-agent allows creating the distributed cognitive systems with a high level of behavior complexity.
It is significant that the cognitive systems (Figure 9), unlike other well-known solutions (CERT/SCIRT, MSSP/MDR, SOC 2.0, IDS/IPS, etc.), have the ability to independently learn and behave in the real conditions of destructive hardware and software of intruders, affecting the protected critical information infrastructure. This will effectively solve the following tasks:
• Recognize patterns (patterns and clusters) that determine the preparation and the beginning of computer aggression;
• Training and development of the typical scenarios of warning, detection and counteraction in cyberspace;
• Generation, accumulation and processing of the new knowledge about the quantitative laws of opposition in cyberspace; • Representations of the “deep” semantics of confrontation in cyberspace; • Preparation and implementation of the adequate decisions, in response to cyber - attack.
The publication was carried out with the financial support of Russian Foundation for Basic Research (RFBR) in the framework of the scientific project No. 20-04-60080.