Adjusting Dynamics of Hopfield Neural Network via Time-Variant Stimulus

Xuenan Peng, Chengqing Li, Yicheng Zeng, Chun-Lai Li

Introduction

Hopfield Neural Networks (HNNs) serve as foundational models for understanding nonlinear dynamics and biological memory. While their ability to exhibit chaotic behavior is well-documented, HNNs are notoriously sensitive to external disturbances, which can either enhance or destroy their chaotic properties. Historically, achieving specific, controllable chaotic outcomes—such as multi-scroll attractors—often required specialized hardware like memristors. This research introduces a method to precisely tune HNN dynamics by injecting time-variant stimuli into the network's weight matrices and state variables, providing a predictable path to complex chaos without hardware redesign.

Problem Statement

The primary challenge in dynamical systems research is dynamic modulation: the ability to predictably expand the chaotic range and complexity of a system using simple, external signals. While previous studies have induced complex "scrolls" (patterns in the system's state space) through internal component modification, there remains a gap in achieving these results through external modulation. The authors address how time-varying signals can be used to generate high-entropy chaotic sequences suitable for high-security applications.

Methodology: The Adjusted Hopfield Neural Network (AHNN)

The researchers developed the Adjusted Hopfield Neural Network (AHNN), modifying the standard mathematical model by introducing three distinct types of stimuli:

  1. Weight Matrix Stimulus (WMS): A time-variant signal that oscillates connection weights between binary states.
  2. State Variable Stimulus (SVS): A signal added directly to the neurons' internal states.
  3. Constant Stimulus (CS): A steady-state signal used to evaluate the limits of system stability.

Hardware Implementation

To validate the theoretical model, the team implemented the AHNN on a Xilinx FPGA using the Runge-Kutta (RK-4) numerical method. A significant hurdle in digital implementation is the computationally expensive hyperbolic tangent ($tanh$) activation function. The authors bypassed this by employing a piecewise Taylor series approximation, allowing for high-fidelity chaotic dynamics with minimal resource overhead.

Cryptographic Application

The AHNN was integrated into an image encryption scheme. The chaotic sequences generated by the network drive a 2D Cat map for pixel diffusion, followed by XOR-based confusion to ensure high security.

Key Insights and Dynamics

Modular Complexity and the "Building Block" Effect

The study revealed that chaotic complexity can be scaled modularly. Deploying four WMSs creates a 4-scroll attractor; adding a single SVS doubles this to an 8-scroll attractor. By integrating multiple SVSs, the system can generate 3D grid-multi-scroll attractors, which significantly increases the randomness available for cryptographic keys.

The Role of Time-Variance

Not all stimuli improve the system. The introduction of a Constant Stimulus (CS) was found to "degenerate" the chaos, eventually forcing the system into a predictable fixed point. This finding underscores that modulation must remain time-variant to maintain the high-entropy chaos required for secure communication.

Results and Performance Analysis

The AHNN-based encryption system was evaluated using standard benchmark images (e.g., lena.tif, pout.tif) and synthetic images to test resistance to statistical attacks.

  • Statistical Randomness: The generated chaotic sequences passed all 15 tests in the NIST SP800-22 suite, confirming they are statistically indistinguishable from true random noise.
  • Information Entropy: The system achieved an entropy value of approximately 7.999 (approaching the theoretical ideal of 8), indicating nearly perfect pixel distribution in encrypted images.
  • Correlation and Robustness: Correlation between adjacent pixels dropped from ~0.9 to ~0.001. Furthermore, the scheme remained effective even when subjected to 10% "salt and pepper" noise or 6% data loss, demonstrating high robustness against transmission errors.

Limitations and Future Directions

While the piecewise Taylor series is efficient for hardware, the authors noted that fitting errors can lead to "dynamic degeneration" over extremely long iterations compared to the continuous theoretical model. Additionally, this study focused on a 3-neuron HNN; scaling this approach to larger networks while maintaining precise control over multi-scroll attractors remains an open research question. Future work may explore fractional-order HNNs and other non-autonomous stimuli to further enhance attractor complexity.

Summary updated

BibTeX
@article{peng2024adjusting,
  author = {Xuenan Peng and Chengqing Li and Yicheng Zeng and Chun-Lai Li},
  title = {{Adjusting Dynamics of Hopfield Neural Network via Time-Variant Stimulus}},
  journal = {IEEE Trans. Circuits Syst. I Regul. Pap.},
  year = {2024},
  doi = {10.1109/tcsi.2024.3390628},
}
Figure