[Paper Review] A deep learning based known plaintext attack method for chaotic cryptosystem

Fusen Wang, Jun Sang+3 more|arXiv (Cornell University)|Mar 9, 2021

Chaos-based Image/Signal Encryption20 references4 citations

TL;DR

This paper proposes a deep learning-based known-plaintext attack (KPA) on chaotic cryptosystems using convolutional neural networks (CNNs) to learn the plaintext-ciphertext mapping. By training CNNs on large datasets of plaintext-ciphertext pairs, the model acts as a decryption system, achieving high-fidelity image reconstruction across multiple chaotic encryption schemes, demonstrating generalization across different systems and networks.

ABSTRACT

In this paper, we propose a known-plaintext attack (KPA) method based on deep learning for traditional chaotic encryption scheme. We employ the convolutional neural network to learn the operation mechanism of chaotic cryptosystem, and accept the trained network as the final decryption system. To evaluate the attack performance of different networks on different chaotic cryptosystem, we adopt two neural networks to perform known-plaintext attacks on two distinct chaotic encryption schemes. The experimental results demonstrate the potential of deep learning-based method for known-plaintext attack against chaotic cryptosystem. Different from the previous known-plaintext attack methods, which were usually limited to a specific chaotic cryptosystem, a neural network can be applied to the cryptanalysis of various chaotic cryptosystems with deep learning-based approach, while several different networks can be designed for the cryptanalysis of chaotic cryptosystems. This paper provides a new idea for the cryptanalysis of chaotic image encryption algorithm.

Motivation & Objective

• To develop a flexible, generalizable known-plaintext attack method for chaotic cryptosystems that overcomes the limitations of traditional, system-specific attacks.
• To investigate whether deep learning models can learn and reverse the complex transformations in chaotic encryption schemes without prior knowledge of the internal parameters.
• To evaluate the performance of different CNN architectures in decrypting various chaotic encryption schemes and compare their reconstruction accuracy and training efficiency.
• To explore the potential of transferring trained models across different chaotic cryptosystems, enhancing attack adaptability.
• To open new research directions in multimedia security by identifying vulnerabilities in chaotic encryption under deep learning-based cryptanalysis.

Proposed method

• Train two distinct convolutional neural networks—Unet and a proposed Multi-Stage Encoder-Decoder Network (MSEDNet)—on large-scale datasets of plaintext-ciphertext pairs generated from two chaotic encryption schemes.
• Use an end-to-end encoder-decoder architecture with skip connections to preserve spatial information and enable pixel-wise regression for image reconstruction.
• Apply standard deep learning components: 3x3 convolutions, batch normalization, ReLU activation, max-pooling (downsampling), and deconvolutional upsampling with concatenation.
• Train the networks using L1 loss to minimize reconstruction error and evaluate performance using correlation coefficient with the original plaintext.
• Employ both MNIST and CIFAR-10 datasets to test generalization across image distributions and evaluate model robustness.
• Use the trained model as a surrogate decryption system, treating it as an equivalent key to recover plaintext from new ciphertexts.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on one chaotic encryption scheme successfully decrypt ciphertexts from a different chaotic cryptosystem?
• RQ2 How does the performance of different CNN architectures (e.g., Unet vs. MSEDNet) compare in reconstructing plaintext from chaotic ciphertexts?
• RQ3 To what extent does the complexity of the encryption mechanism affect the reconstruction accuracy of deep learning-based KPA?
• RQ4 Can a single trained neural network generalize across multiple chaotic encryption schemes, or is retraining required for each system?
• RQ5 How do training time, model capacity, and network depth influence the success rate and fidelity of the decryption process?

Key findings

• The proposed deep learning-based KPA achieved a testing accuracy of 99.15% with Unet and 97.80% with MSEDNet on the Logistic map encryption scheme, demonstrating high reconstruction fidelity.
• On the more complex Hybrid Chaotic map (CIFAR-10), Unet achieved 95.70% testing accuracy, while MSEDNet reached 93.62%, showing that network architecture significantly affects performance.
• Unet outperformed MSEDNet in both training efficiency and accuracy across both datasets, with lower training loss and higher correlation coefficients.
• The results confirm that different neural networks yield varying decryption effects, indicating that model choice directly impacts reconstruction quality.
• The same CNN architecture (e.g., Unet) successfully attacked multiple distinct chaotic cryptosystems, proving the method’s generalization potential beyond single-system targeting.
• The attack method generalizes across image datasets (MNIST and CIFAR-10), indicating robustness to input distribution shifts in real-world scenarios.