Cryptographic analysis on an optical random-phase-encoding cryptosystem for complex targets based on physics-informed...

Area of Science:

• Optics and Information Security
• Applied Deep Learning
• Cryptography

Background:

• Deep learning (DL) methods in optical cryptanalysis often lack physical priors, limiting their practical application and generalization.
• Existing methods struggle with generalization and are constrained by image size limitations.

Purpose of the Study:

• To demonstrate that double-random phase encoding (DRPE) optical cryptosystems are vulnerable to a preprocessing ciphertext-only attack (pCOA) using DL strategies.
• To enhance the generalization capabilities of DL-based optical cryptanalysis by integrating physical knowledge.
• To overcome the image size limitations of traditional ciphertext-only attack (COA) methods.

Main Methods:

• A preprocessing ciphertext-only attack (pCOA) based on DL strategies is proposed for DRPE optical cryptosystems.
• A physical knowledge deep learning method is employed, incorporating physical priors to learn statistical invariants from preprocessed ciphertext.
• The method utilizes one random phase mask (RPM) for initial training and scales with an increasing number of training RPMs.

Main Results:

• High prediction fidelity for complex targets is achieved, even with a single RPM for training.
• The generalization ability of the DL model is significantly improved by increasing the number of training RPMs.
• The proposed method successfully breaks the image size limitations inherent in traditional COA methods.

Conclusions:

• The developed learning-based pCOA method is feasible and effective for optical cryptanalysis.
• Integrating physical priors into DL models substantially enhances their generalization capabilities in optical security applications.
• This approach offers a more robust and versatile solution for breaking DRPE-based optical encryption.