Reservoir computing-based cryptanalysis of structured phase-masked chaos encryption

Area of Science:

• Cryptography
• Machine Learning
• Signal Processing

Background:

• Chaos-based encryption offers potential for secure data transmission.
• Learning-assisted cryptanalysis poses a significant threat to existing encryption schemes.
• Echo State Networks (ESNs) are effective for time-series analysis and pattern recognition.

Purpose of the Study:

• To evaluate the robustness of a specific chaos-based audio encryption scheme against cryptanalysis.
• To investigate the effectiveness of Echo State Networks (ESNs) in breaking this encryption under partial leakage conditions.
• To analyze the impact of varying levels of leakage on decryption success.

Main Methods:

• A chaos-based audio encryption pipeline utilizing the Rössler attractor for key stream generation.
• Implementation of a partial-leakage threat model with random missing leakage for phase information.
• Training an ESN on observed leakage blocks to impute missing phase data and estimate the permutation and phase mask.
• Evaluation of reconstruction quality using Signal-to-Noise Ratio (SNR), STOI, CORR, and Normalized Mean Square Error (NMSE).

Main Results:

• The study demonstrates a non-linear relationship between the amount of permutation leakage and the success of adversarial recovery.
• ESNs can effectively impute missing phase information, enabling approximate decryption.
• The effectiveness of the attack is influenced by frequency-bin alignment and the learnability of the phase model.

Conclusions:

• The investigated chaos-based audio encryption scheme exhibits vulnerabilities to learning-assisted cryptanalysis, particularly under partial leakage scenarios.
• The findings highlight the critical interplay between leakage amount, frequency-bin alignment, and phase-model learnability in determining attack success.
• Further research is needed to develop more resilient chaos-based encryption methods against advanced cryptanalytic techniques.