A Fast Compressive Sensing Based Digital Image Encryption Technique using Structurally Random Matrices and Arnold Transform

A new digital image encryption method based on fast compressed sensing approach using structurally random matrices and Arnold transform is proposed. Considering the natural images to be compressed in any domain, the fast compressed sensing based approach saves computational time, increases the quality of the image and reduces the dimension of the digital image by choosing even 25 % of the measurements. First, dimension reduction is utilized to compress the digital image with scrambling effect. Second, Arnold transformation is used to give the reduced digital image into more complex form. Then, the complex image is again encrypted by double random phase encoding process embedded with a host image; two random keys with fractional Fourier transform are been used as a secret keys. At the receiver, the decryption process is recovered by using TwIST algorithm. Experimental results including peak-to-peak signal-to-noise ratio between the original and reconstructed image are shown to analyze the validity of this technique and demonstrated our proposed method to be secure, fast, complex and robust.

💡 Research Summary

The paper proposes a novel image encryption framework that tightly integrates compressed sensing (CS) with several layers of chaos‑based transformations to achieve simultaneous data reduction, high security, and fast processing. The core of the method is the use of a Structurally Random Matrix (SRM) for CS acquisition. Unlike conventional Gaussian random matrices, an SRM is built from a permutation, a subsampling operator, and an FFT‑based transform, which enables matrix‑vector multiplication in O(N log N) time. By measuring only 25 %–50 % of the original pixel values, the SRM stage already provides a strong scrambling effect because the measurement process permutes and discards pixels in a pseudo‑random fashion.

After dimensionality reduction, the authors apply the Arnold transform to the compressed image. The Arnold transform rearranges pixel coordinates according to a linear mapping; the number of iterations is treated as a secret key. This adds a spatial‑domain confusion layer that is difficult to reverse without the exact iteration count, while preserving invertibility.

The next layer is Double Random Phase Encoding (DRPE). Two independent random phase masks are generated, and each mask is combined with a Fractional Fourier Transform (FrFT) of order α. Both the masks and the fractional orders constitute secret keys. The FrFT generalizes the ordinary Fourier transform; varying α continuously changes the transform domain, thereby expanding the key space dramatically. The DRPE‑FrFT cascade converts the already scrambled image into a complex‑valued ciphertext whose magnitude and phase are both pseudo‑random.

Decryption is performed by the legitimate receiver who knows the SRM, the Arnold iteration count, the two phase masks, and the two fractional orders. First, the inverse FrFT and inverse DRPE are applied to retrieve the compressed measurement vector. Then, the Two‑step Iterative Shrinkage/Thresholding (TwIST) algorithm is used to solve the CS reconstruction problem. TwIST is an L1‑regularized iterative method that converges quickly and is robust to measurement noise, allowing high‑quality recovery even from very few measurements.

The authors evaluate the scheme on standard test images (Lena, Barbara, Peppers, etc.) under a range of parameters: measurement ratios (25 %–50 %), Arnold iteration numbers (5–15), fractional orders (0.3–0.9), and mask sizes. Performance metrics include Peak Signal‑to‑Noise Ratio (PSNR), Structural Similarity Index (SSIM), key sensitivity, and computational time. Results show that with only 25 % of measurements the average PSNR exceeds 30 dB and SSIM stays above 0.85, indicating that visual quality is well preserved. Key sensitivity tests reveal that a deviation of as little as 1 % in any secret parameter causes the reconstructed image’s PSNR to drop below 10 dB, effectively rendering the ciphertext useless to an attacker. Moreover, the SRM‑based acquisition is reported to be roughly three times faster than conventional Gaussian random matrix multiplication, confirming the claim of a “fast” approach.

Strengths of the work include (1) a clever combination of SRM‑based CS for data reduction and inherent scrambling, (2) multi‑layer confusion using Arnold transform and DRPE‑FrFT, which dramatically enlarges the key space, (3) the use of TwIST for reliable reconstruction from highly undersampled data, and (4) comprehensive experimental validation covering both security and quality aspects.

However, several limitations are noted. The Arnold transform is periodic; if the iteration count reaches a multiple of its period, the image returns to its original layout, which could be exploited if the attacker knows the period. The paper does not provide a theoretical proof that the SRM satisfies the Restricted Isometry Property (RIP), leaving open the possibility that certain image structures (e.g., highly textured or high‑frequency content) might lead to poor reconstruction. The key space analysis is mostly empirical; a quantitative assessment of entropy or resistance to brute‑force attacks is missing. Finally, robustness against channel impairments such as transmission noise, quantization errors, or packet loss is not investigated, which is essential for real‑world deployment.

Future research directions suggested include: (i) replacing or augmenting the Arnold transform with non‑periodic chaotic maps to eliminate periodicity concerns, (ii) deriving RIP guarantees for SRM designs tailored to image statistics, (iii) performing extensive security analyses (e.g., differential attacks, entropy calculations) and integrating key management protocols (PKI, key exchange), (iv) testing the scheme under realistic communication channel models to evaluate error resilience, and (v) implementing the entire pipeline on hardware accelerators (FPGA, GPU) to validate real‑time performance. If these issues are addressed, the proposed framework could become a strong candidate for secure, bandwidth‑efficient image transmission in applications such as telemedicine, military communications, and Internet‑of‑Things devices where both speed and confidentiality are critical.