As the cornerstone of network information transmission, the optical fiber communication network currently carries more than 90% of the global data traffic transmission, and its security is vital to maintaining information transmission privacy. Optical fiber communication is traditionally considered to be relatively secure. However, with the progress of communication technology, such as the application of wavelength division multiplexing and optical amplification technology, despite the greatly improved transmission capacity and distance, concerns about the potential eavesdropping risk of optical fiber communication systems are caused. At present, secure communication technology is mainly divided into two categories of mathematical algorithm encryption and physical layer encryption, with more attention paid to the latter category because of its ability to face high performance computing threats. Meanwhile, quantum key distribution provides absolute security in theory, but faces technical obstacles in implementation, such as low efficiency of single photon detection and large transmission loss, thus limiting its practicability. As a method of physical layer encryption, chaotic secure communication employs the randomness of chaotic optical signals to encrypt information. However, it is a big challenge to realize wide-band chaotic synchronization in high-speed systems, mainly because the initial value sensitivity of chaotic systems makes it difficult to accurately match the parameters of the receiving and sending terminals. To solve the difficulty of chaos synchronization at the receiving and sending ends of traditional chaotic encryption communication systems, some studies have proposed to adopt deep learning technology to realize chaos generation and synchronization, but most of the current studies are only offline processing on computers. Therefore, we propose a method based on deep learning technology to model the optoelectronic oscillation chaotic source to realize the digital domain generation of chaos. Additionally, the chaotic AI model trained by the host computer is deployed on the field programmable gate array (FPGA) and applied to real-time chaotic sequence generation and random number generation.

Based on the long short-term memory (LSTM) network, the optoelectronic oscillation chaotic source is modeled. After pruning optimization, the model is deployed on FPGA to generate chaotic sequences and the random number in real time by a digital-to-analog converter (DAC) chip. Firstly, the chaotic AI model trained by the host computer is optimized and pruned, and the effective network parameters are imported into FPGA and saved. Then the hardware structure of the chaotic AI model is designed on FPGA to generate chaotic sequences. Meanwhile, the chaotic AI model is utilized for real-time random number generation based on the post-processing method optimized by the least significant bit selection. Finally, the chaotic sequence and random number generated by the model are converted into analog signal output by the DAC chip.

Based on the chaotic AI model deployed on FPGA, the DAC chip is driven to output chaotic waveform in real time at a sampling frequency of 70 MHz. Compared with the chaotic AI model before optimization, the FPGA’s DSP module resource consumption is reduced by about 31.7%, the Block RAM (BRAM) resource consumption is reduced by 58%, the model calculation delay is reduced by about 44.4% (Table 2), and the model’s prediction accuracy has almost no loss (Fig. 9). By drawing the phase diagram of the original optoelectronic oscillation chaotic source and the chaotic AI model deployed on the FPGA (Fig. 10), it is proved that the deployed model retains the chaotic output characteristics of the original optoelectronic oscillation (OEO) system. By adding a small disturbance of order 10^-8 to the excitation signal of the chaotic AI model, the mean square error of the output of the chaotic AI model before and after adding the perturbation is calculated (Fig. 11), which verifies the sensitivity of the model to the initial value and further proves the output chaos of the model. Based on the post-processing method optimized for the least significant bit selection, random numbers are generated at a rate of 70 Mbit/s, and the results can pass the NIST SP 800-22 randomness test (Table 3).

To solve the problem of wide-band chaos synchronization in traditional chaotic encryption communication systems, we model the optoelectronic oscillation chaotic source based on the LSTM network. After optimized pruning, the chaotic AI model is deployed on FPGA, and the DAC chip is driven to output a chaotic sequence in real time at a sampling frequency of 70 MHz. Compared to pre-optimization, the DSP module resource footprint on the FPGA is reduced by about 31.7%, the BRAM resource footprint is reduced by 58%, and the computation latency is reduced by about 44.4%. The chaotic output characteristic of the deployed model maintaining the original OEO system is proved by the phase diagram. The sensitivity of the model to the initial value is verified by calculating the mean square error between the output of the chaotic AI model before and after adding small perturbations. Additionally, based on the least significant bit selection optimization post-processing method, we further apply the deployed chaotic AI model to generate a real-time random number with 70 Mbit/s bit rate, and the obtained results can pass the NIST SP 800-22 test.