Research progress in cryptography intelligence
In the new era of the Internet of Everything opened by the rapid development of artificial intelligence, 5G and network technology, people have put forward higher requirements for secure communication and data security.
With the improvement of computer computing power and the development of quantum computers, many cryptographic algorithms popular in the 19th century were cracked, and traditional cryptography faced severe challenges.
In recent years, the development of cryptography has entered the stage of intelligence, and the research on cryptography intelligence using machine learning technology has begun to enter people's field of vision.
The birth of the Intuitive Cryptography algorithm and Intuitive Cryptography Analysis is the main feature of cryptography entering the intelligent stage.
Intuitive Cryptography algorithms design, analyze, and implement encryption and decryption based on machine learning technology, or use the synchronization mechanism of neural networks to achieve key exchange.
Not only that, such algorithms can also realize on-demand encryption by reasoning the secret information part of the machine learning model, taking into account the efficiency and security of encryption, and have the advantages of high flexibility and strong adaptive ability.
Intelligent Cryptanalysis is based on machine learning technology to mine features in secret information, and combines feature extraction with cryptographic research to conduct cryptanalysis more efficiently and accurately.
Machine learning and cryptography have a lot in common, such as the ability to handle large amounts of data and a large search space.
Back in 1991, Ro-nald Rivest, one of the founders of RSA, specialized in discussing the similarities and differences between machine learning and cryptography, and the interaction between the two fields.
Today, researchers at home and abroad have made a series of achievements in this field, and initially formed a new crossover field - neural network cryptography.
As a large branch of machine learning, artificial neural networks are widely used in the design of Intuitive Password algorithms.
Since the nonlinear mapping characteristics of artificial neural networks coincide with the nonlinear mapping design principles of cryptographic algorithms, the use of artificial neural networks to carry out intelligent cryptographic algorithm research has become an important branch of the development of modern cryptography.
At present, according to the working mechanism of artificial neural networks, the artificial neural networks used in intelligent cryptographic algorithms can be divided into feedback networks (such as TPM neural networks, etc.) and feed-forward networks (such as GAN, chaotic neural networks, etc.):
1) Feedback neural network is a large-scale nonlinear dynamic system, in the intelligent cryptography algorithm can use the mutual learning of neural networks to derive the shared secret, this kind of intelligent cryptography algorithm does not need to transmit and distribute the key, similar to the key agreement protocol;
2) Feedforward neural networks can be mathematically seen as a large-scale nonlinear mapping system that can be used in the design of symmetric and asymmetric cryptosystems.
GAN can be used to automatically design and generate cryptographic systems, that is, AI can be used to generate cryptographic systems. Using chaotic neural networks can design more chaotic and diffuse cryptosystems.
One. TPM-based Intuitive Password algorithm
The earliest research on TPM dates back to the work of Kinzel et al. and Rosen et al. in 2002. They analyzed the TPM mutual learning problem and proposed for the first time to use the TPM mutual learning mechanism to construct a key exchange protocol, and the receiving side only needs to perform a limited number of input-output exchange steps to converge to the synchronous state.
In terms of security analysis of TPM, Rosen et al. believe that attackers cannot discover keys by training another neural network, or affect the mutual learning process of TPM through flipping attacks, so as to ensure the security of communication.
Klein et al. used statistical physics to analyze the synchronization process of TPM mutual learning, and tested the security under different attack strategies, and found that the communicator with two-way learning was more likely to synchronize the neural network than the attacker with one-way learning.
Compared with cryptographic algorithms based on number theory, TPM-based Intuitive Cryptography algorithms have 3 advantages:
1) The operation is simple, and its training algorithm is essentially a linear filter, which is easy to implement in hardware;
2) The number of calculations to generate keys is small, and it only takes N calculations to generate keys with length N;
3) For each communication, a new key can be generated for each block of the message.
2. GAN-based Intuitive Cryptography algorithm
Based on the iterative principle of generative adversarial of GAN, researchers propose many new ideas for intelligent cryptographic algorithms.
Abadi et al. show how to protect the communication process by using GAN without specifying a cryptographic algorithm and related conditions.
The principle of the GAN-based cryptographic algorithm is shown in Figure 2. The advantages of GAN-based Intuitive Password algorithms are:
1) The GAN model in the existing symmetric cryptosystem uses the backpropagation of the neural network, and there is no need to infer hidden variables during training;
2) The improvement of the ability to fight the attacker's deciphering ability will enhance the robustness of the generated cryptographic algorithm.
3. Intelligent cryptography algorithm based on chaotic neural network
The researchers used chaotic neural networks to explore a series of new intelligent cryptographic algorithms.
Zhang et al. discuss the connection between cryptography and chaos theory, and train neural networks using the initial sensitivity of chaos to implement encryption and decryption algorithms.
Su et al. proposed a chaotic neural network model for digital signal encryption and decryption and its VLSI structure, which uses the binary sequence generated by the chaotic system to set the bias and weight of neurons to encrypt digital signals.
The advantages of Intuitive Cryptography algorithms based on chaotic neural networks are:
1) The asynchronous encryption model is practical and simple, and does not require atomic operations;
2) Fast encryption speed, no distortion, suitable for system integration;
3) HIGH NONLINEAR COMPLEXITY, GOOD PARALLELISM, MEET THE REQUIREMENTS OF IPV6 CONFIDENTIAL COMMUNICATION.
Through the above review, the combination of cryptography research and machine learning will become an important branch of cryptography development under the background of rapid improvement of computer computing power and breakthrough in the application of deep learning technology.
