Cryptanalysis of Dual-Stage Permutation Encryption Using Large-Kernel Convolutional Neural Network and Known Plaintext Attack

Ching-Chun Chang, Shuying Xu, Kai Gao, Chin-Chen ChangCryptogr 2024 · DOI

TL;DR

This paper cracks a dual-stage image encryption scheme designed for reversible data hiding. The authors demonstrate two attack methods: a known-plaintext attack that exploits pixel value preservation and a novel approach using a large-kernel convolutional neural network (LKCNN) to identify encryption parameters without a plaintext. Their findings reveal significant security vulnerabilities, questioning the scheme's suitability for secure cloud storage.

What Problem Does This Paper Address?

Cloud storage offers convenience but raises serious security concerns for sensitive data like images. Reversible data-hiding in encrypted images (RDHEI) is a promising solution, allowing data to be hidden within encrypted images before uploading to the cloud. This requires robust encryption that preserves enough image structure for data hiding. A recent dual-stage permutation encryption scheme was proposed for RDHEI, claiming high embedding capacity and security. However, this paper questions its actual security, arguing that the preservation of pixel value distribution, while beneficial for data hiding, can be exploited by attackers. The core problem addressed is the security of this specific dual-stage permutation encryption scheme and the development of effective cryptanalysis techniques to expose its weaknesses.

Methods

The paper proposes two main cryptanalysis methods:

  • Known-Plaintext Attack (KPA) with Block Modulation: This method assumes the attacker has access to at least one original plaintext image and its corresponding ciphertext. The core idea is to exploit the fact that the dual-stage permutation encryption preserves the pixel value distribution. The attacker divides both plaintext and ciphertext into blocks and sorts the pixels within each block in descending order. By comparing these "modulated" blocks, the attacker can estimate the inner-block shuffling sequence and the coarse-level block permutation sequence. This allows for the reconstruction of the original image.
  • Large-Kernel Convolutional Neural Network (LKCNN) for Parameter Estimation: This method is designed for scenarios where the attacker does not have a plaintext image. It leverages the observation that the mean pixel value of each 2x2 block remains unchanged after encryption. The attacker constructs a "mean sequence" from the encrypted image. By iteratively applying random permutation sequences (generated from a limited key space) and comparing the resulting mean sequences to the original image's mean sequence, the attacker aims to find the correct permutation. A large-kernel convolutional neural network (LKCNN) is trained to automatically detect the correct permutation sequence by classifying whether a given sequence aligns with the original image's characteristics.

Datasets

  • Training and Testing: The "Bossbase" dataset was used to construct training and testing samples for the LKCNN. Images were downscaled to 256x256 pixels.
  • Positive Samples: Created from original images by dividing them into 2x2 blocks and calculating their mean values.
  • Negative Samples: Created by first applying the dual-stage permutation encryption to the downscaled images and then calculating the mean values of the resulting blocks.
  • Evaluation: Four standard grayscale test images ("F16", "Barbara", "Peppers", and "Baboon") were used for evaluating the cryptanalysis methods. These images were resized to 256x256 for testing.

Key Insights and Ideas

  • Pixel Value Preservation is a Double-Edged Sword: The very property that makes the encryption scheme suitable for reversible data hiding (preserving pixel value distribution and intra-block correlation) is also its primary vulnerability. Attackers can exploit these statistical properties to infer encryption parameters.
  • LKCNN for Automated Parameter Discovery: The introduction of an LKCNN to automatically identify the correct permutation sequence without a known plaintext is a novel approach. This bypasses the need for exhaustive key searching in certain scenarios and leverages deep learning for cryptanalysis.
  • Limited Key Space is Exploitable: The paper highlights that even with a large theoretical number of permutations, the practical limitations of random number generators and finite key spaces in hardware make exhaustive search feasible for attackers.
  • Visual Quality vs. Information Leakage: The KPA demonstrates that even with moderate estimation accuracy, significant portions of the original image's information can be leaked, leading to recognizable decrypted content, especially for smoother images.

Key Results

  • KPA Effectiveness: The KPA achieved high estimation accuracy for both the block permutation sequence (around 97% for "Baboon") and the inner-block shuffling sequence (around 95% for "Baboon") with a single known plaintext. With two known plaintexts, accuracy approached 100%.
  • LKCNN Accuracy: The LKCNN achieved high classification accuracy: 92.23% for positive samples and 100% for negative samples on a test dataset. Misclassifications were primarily observed in complex image details.
  • Reconstruction Quality: Reconstructed images using the proposed methods showed significantly improved PSNR (around 20 dB) and SSIM compared to a prior state-of-the-art method (around 11-15 dB PSNR).
  • Computational Efficiency: The KPA's average running time was approximately 362.61 seconds per image. The LKCNN approach was also computationally efficient, with sequence acquisition taking 0.8s and evaluation taking 0.008s.

Limitations and Future Work

  • Image Resizing: The scheme resizes images to 256x256 pixels for practical reasons related to hardware and processing time. The authors suggest that maintaining original image sizes would be more optimal if sufficient resources were available.
  • Key Space Limitations: While the LKCNN operates within a limited key space, attempting all possible keys can still be computationally challenging due to hardware limitations.
  • LKCNN Misclassifications: The LKCNN occasionally misclassifies samples, particularly those with complex details, which can affect reliability and potentially lead to inaccurate interpretations.
  • Future Work: The authors suggest further investigation into the security implications of dual-stage permutation encryption and the development of more resilient techniques to mitigate identified vulnerabilities. They also note that improving LKCNN's accuracy on complex details is an area for future research.

Summary updated

BibTeX
@article{chang2024cryptanalysis,
  author = {Ching-Chun Chang and Shuying Xu and Kai Gao and Chin-Chen Chang},
  title = {{Cryptanalysis of Dual-Stage Permutation Encryption Using Large-Kernel Convolutional Neural Network and Known Plaintext Attack}},
  journal = {Cryptogr.},
  year = {2024},
  doi = {10.3390/cryptography8030041},
}
Figure