An Online Evaluation Method for Random Number Entropy Sources Based on Time-Frequency Feature Fusion

Area of Science:

• Cryptography and Information Security
• Machine Learning and Artificial Intelligence
• Hardware Security

Background:

• Traditional entropy source evaluation methods, often statistical, lack on-chip or online deployment capabilities.
• Online assessment of entropy source quality is crucial for high-level encryption applications.
• Existing methods struggle to provide real-time feedback on the randomness quality.

Purpose of the Study:

• To establish a novel, online-deployable method for assessing entropy source quality.
• To investigate the correlation between minimum entropy and prediction accuracy of random sequences.
• To develop an efficient deep learning model for real-time entropy source evaluation.

Main Methods:

• Experimental analysis correlating minimum entropy values with prediction accuracy using Pearson correlation coefficient.
• Development and application of a novel deep learning architecture: Fast Fourier Transform-Attention Mechanism-Long Short-Term Memory Network (FFT-ATT-LSTM).
• Integration of Fast Fourier Transform (FFT) with a simplified soft attention mechanism for feature fusion.

Main Results:

• A significant negative correlation (r = -0.925, p < 1.07 × 10^-7) was found between minimum entropy and prediction accuracy.
• The proposed FFT-ATT-LSTM model achieved improved prediction accuracy by 4.46% and 8% over baseline networks.
• FFT-ATT-LSTM demonstrated a compact parameter size (33.90 KB), outperforming TCN and Transformers in efficiency.

Conclusions:

• Minimum entropy serves as a reliable indicator for predicting random sequence quality.
• The FFT-ATT-LSTM model provides an accurate and resource-efficient solution for online entropy source evaluation.
• The developed model has significant potential for practical application in secure systems requiring real-time randomness assessment.