• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Sunday, November 22, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

Deep Learning Steganography Framework Combines Autoencoders and Cryptography to Secure Digital Images

by
November 21, 2026
in Technology
Reading Time: 6 mins read
0
Deep Learning Steganography Framework Combines Autoencoders and Cryptography to Secure Digital Images

Deep Learning Steganography Framework Combines Autoencoders and Cryptography to Secure Digital Images

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Securing sensitive visual communications across global digital networks requires robust methods that conceal both the message content and its very transmission. In an investigative paper published in the Journal of Umm Al-Qura University for Engineering and Architecture, researchers Ahd Aljarf and Yasmeen Rashidi introduced DL-Steg, a hybrid deep learning steganography framework designed to hide secret images within digital cover containers while maintaining visual imperceptibility and cryptographic protection. Classical information hiding methodologies, including spatial-domain least significant bit substitution, transform-domain discrete cosine transforms, and discrete wavelet transforms, frequently struggle to maintain a functional equilibrium between container distortion, data hiding capacity, and resistance to forensic steganalysis. When embedding volumes increase, classical systems introduce noticeable statistical artifacts that expose hidden transmissions to passive and active steganalysis classifiers. To resolve these vulnerabilities, the researchers developed an architecture that coordinates three distinct computational mechanisms: preprocessing transformations, asymmetric cryptographic encryption, and a dual-stage deep learning pipeline utilizing stacked autoencoders and recurrent neural networks.

The conceptual boundary separating cryptography from steganography forms the foundational justification for the DL-Steg design. While conventional cryptographic protocols scramble visual information into ciphertext to prevent unauthorized interpretation, they overtly signal the existence of sensitive material to network observers. Steganography embeds secret data directly inside unsuspicious host media such as digital photographs, aiming to render the embedded artifact completely invisible to human inspection and statistical inspection algorithms. To eliminate single-point vulnerabilities where intercepted stego images could yield plaintext payloads if computational steganalysis succeeds, DL-Steg introduces a dual-layer defensive posture. Prior to spatial embedding, secret imagery undergoes mathematical transformation through Elliptic Curve Cryptography using the standardized brainpoolP256r1 prime field Weierstrass curve. By multiplying ephemeral private keys against recipient public keys across 256-bit prime coordinates, the system derives shared secret points that are hashed and combined with numerical image blocks through bitwise XOR operations, ensuring that the payload remains mathematically unreadable even if an adversary successfully isolates the embedded data layer.

To train, evaluate, and calibrate the steganographic network under standardized visual conditions, the investigators utilized the Classified ImageNet dataset sourced from the Kaggle repository. This dataset comprises 540,000 high-resolution images categorized across 1,000 distinct thematic classes spanning fauna, flora, vehicles, and manufactured objects. Experimental protocols partitioned the image collection into an eighty percent training allocation and a twenty percent testing split. Input images exhibited diverse dimensional scales and therefore required structured preprocessing before ingestion by the neural layers. Every candidate image was rescaled to a uniform spatial resolution of 256 by 256 pixels, followed by floating-point normalization achieved by dividing absolute pixel values by 255. Subsequent statistical standardization subtracted feature-level mean vectors across dimensions and scaled variance against standard deviation metrics, aligning the visual distribution with Zero Component Analysis principles and preparing both cover containers and secret targets for coordinated feature transformation.

The dimensional compression and core embedding within DL-Steg are handled by a Stacked Autoencoder configured with paired encoding and decoding stages. Autoencoders function by mapping high-dimensional image matrices into constrained, lower-dimensional bottleneck latent spaces, discarding redundant background data while preserving critical structural features before mapping representations back into reconstruction space. The DL-Steg architecture incorporates two sequential encoding layers operating alongside two corresponding decoding layers, utilizing rectified linear unit and sigmoid activation functions to govern continuous feature representations. In standard deep learning steganography, standard convolutional autoencoders frequently struggle to maintain long-range spatial correlations across distributed pixel regions, resulting in localized visual distortions when secret features merge into host images. DL-Steg circumvents this bottleneck by coupling the latent representations produced by the stacked autoencoder directly into a Long Short-Term Memory network tailored for sequential dependency processing.

Integrating Long Short-Term Memory networks allows the DL-Steg architecture to treat two-dimensional spatial structures as structured temporal sequences, overcoming the spatial range limitations inherent in local convolutional filters. The system reshapes compressed feature maps derived from the stacked autoencoder into ordered sequences of localized pixel patches and vectors. The recurrent memory cells within the Long Short-Term Memory framework regulate information propagation across input, output, and forget gates, governed by hyperbolic tangent and sigmoid mathematical transformations. By calculating element-wise products across weight matrices and historical context vectors, the recurrent network tracks relational dependencies between spatially separated regions across the entire image canvas. This sequential tracking enforces a uniform distribution of embedded features across the cover matrix, eliminating clustered pixel modifications that conventional machine learning steganalysis detectors exploit to identify steganographic media.

The computational workflow was constructed using Python version 3.6.0 alongside the TensorFlow and Keras application programming interfaces, executing on Google Colaboratory workstations equipped with an NVIDIA GeForce GTX 1050 graphical processing unit featuring four gigabytes of dedicated video memory, sixteen gigabytes of system memory, and hybrid storage arrays. Following spatial encoding and Long Short-Term Memory modeling, the synthesized secret representation is merged with the standardized cover image to generate the final container image suitable for digital distribution. Extraction reverses these coordinated operations: the receiving system separates the encoded latent secret data from the carrier container, deploys the stacked autoencoder decoding layers to reconstruct the compressed matrices, and applies reciprocal Elliptic Curve Cryptography private keys to decrypt the numerical blocks back into the original 256 by 256 pixel secret image.

Empirical performance verification assessed the system across standardized visual fidelity metrics, including Peak Signal-to-Noise Ratio, Structural Similarity Index Measure, Mean Squared Error, and absolute payload capacity measured in bits per pixel. Testing five representative cover-secret image pairings demonstrated consistent perceptual fidelity across varied contextual subjects. When embedding a binocular secret image inside a bald eagle cover image, the model achieved a 43.9849 decibel Peak Signal-to-Noise Ratio and an index similarity of 0.9857. An elephant payload embedded into a backpack cover recorded 44.2132 decibels and 0.9735 similarity, while a butterfly embedded in a guitar cover yielded 44.1518 decibels and 0.9760 similarity. An acorn image hidden within a beagle container registered 42.3709 decibels and 0.9707 similarity. The highest single experimental score occurred during the integration of an aircraft carrier secret image into a chameleon cover image, generating a peak ratio of 45.5940 decibels, a 0.9877 Structural Similarity Index, an optimal Mean Squared Error of 0.143, and an information payload capacity of 24 bits per pixel.

To verify that these observed metrics reflected broader algorithmic consistency rather than anomalous image pairings, the researchers conducted aggregated statistical assessments across randomly sampled image distributions drawn from the Classified ImageNet repository. Across this generalized evaluation cohort, DL-Steg sustained a mean Peak Signal-to-Noise Ratio of 45.21 decibels with a standard deviation of 0.67 decibels, bounded within a ninety-five percent confidence interval ranging from 45.08 to 45.34 decibels. Structural similarity across the evaluation cohort averaged 0.9854 with a standard deviation of 0.0041, falling within a confidence interval between 0.9848 and 0.9860. The Mean Squared Error averaged 0.152 with a standard deviation of 0.013 and a confidence interval between 0.148 and 0.156. Execution latency evaluations confirmed operational feasibility, showing an average encryption duration of 1.4 seconds and an average decryption latency of 1.2 seconds, with optimal recorded operational timings reaching 1.4206 seconds for encryption and 1.1891 seconds for extraction decryption.

Comparative evaluations benchmarked DL-Steg against published steganographic architectures reported in recent literature, confirming consistent performance advantages in visual reconstruction and error minimization. The peak signal ratio of 45.5940 decibels represented an improvement ranging from 1.26 percent to 16.50 percent over competing systems, including SegNet-DNN and Ste-Guz architectures, while substantially exceeding the performance of UNet++ encoder variants. In structural similarity, the model achieved a 98.77 percent fidelity rating, surpassing alternative models by margins spanning 0.02 percent to 14.77 percent, where standalone autoencoder baselines degraded to 0.84. To isolate the individual contributions of each architectural module, the authors performed ablation experiments comparing the complete pipeline against three simplified baselines: an isolated autoencoder, an autoencoder combined with recurrent memory layers without cryptography, and an autoencoder paired with cryptography absent recurrent memory. The standalone autoencoder attained only 41.82 decibels with an error rate of 0.226, while adding the recurrent network raised fidelity to 43.95 decibels by capturing sequential dependencies. Coupling cryptography without recurrent memory produced 42.77 decibels, confirming that the full combination of all three modules is required to simultaneously optimize concealment, robustness, and payload capacity.

Despite these empirical advantages, the researchers delineated specific technical constraints and operational overheads associated with the DL-Steg architecture. The multi-tiered combination of stacked autoencoders, recurrent gating networks, and elliptic curve cryptographic operations inherently escalates computational complexity, resulting in extended training durations and demanding hardware resource requirements. Furthermore, achieving optimal reconstruction fidelity requires meticulous hyperparameter tuning when adapting the framework to differing image domains, posing optimization challenges across resource-constrained deployment environments. Looking toward future research directions, the authors emphasized the need to explore advanced optimization algorithms and adaptive learning rate strategies to curtail computational requirements and training overhead. Additionally, future investigations will examine alternate deep neural configurations and lightweight hybrid frameworks to determine whether high-capacity steganographic concealment can be maintained across broader operational settings without sacrificing encryption strength or visual imperceptibility.

Subject of Research: Deep learning image steganography and cryptographic data hiding

Article Title: DL-Steg: a deep learning-based steganography model for improving image security

Article References: Aljarf, A., & Rashidi, Y. (2026). DL-Steg: a deep learning-based steganography model for improving image security. Journal of Umm Al-Qura University for Engineering and Architecture, 17(4), 841-858. https://doi.org/10.1007/s43995-025-00266-2

Image Credits: AI Generated

DOI: 10.1007/s43995-025-00266-2

Keywords: image steganography, deep learning, stacked autoencoder, long short-term memory, elliptic curve cryptography, information hiding, image security, steganalysis resistance, ImageNet dataset, peak signal-to-noise ratio, structural similarity index, mean squared error, payload capacity, brainpoolP256r1, data concealment, recurrent neural networks, visual imperceptibility, lossless transmission

News Source: Blake Davidson. (November 21, 2026). Deep Learning Steganography Framework Combines Autoencoders and Cryptography to Secure Digital Images. Scienmag.

Tags: brainpoolP256r1data concealmentdeep learningElliptic Curve Cryptographyimage securityimage steganographyImageNet datasetinformation hidinglong short-term memorylossless transmissionmean squared errorPayload Capacitypeak signal-to-noise ratiorecurrent neural networksstacked autoencodersteganalysis resistancestructural similarity indexvisual imperceptibility
Share12Tweet7Share2ShareShareShare1

Related Posts

Radio Tags Alter Flight Kinematics in Neotropical Butterflies Below Standard Weight Limits

Radio Tags Alter Flight Kinematics in Neotropical Butterflies Below Standard Weight Limits

November 22, 2026
Strut-and-Tie Model Accurately Predicts Flexural Capacity of Prefabricated Bridge Wet Joints

Strut-and-Tie Model Accurately Predicts Flexural Capacity of Prefabricated Bridge Wet Joints

November 22, 2026

Precast Concrete Shear Walls with UHPC Boundary Connections Improve Seismic Performance

November 21, 2026

Portable Raman Spectroscopy Screens Microbial Carotenoids for Solar Cell Fabrication

November 21, 2026

POPULAR NEWS

  • DCAF15 and ANKIB1 Form Enzyme Complex Driving Inflammatory NLRP3 Activation

    DCAF15 and ANKIB1 Form Enzyme Complex Driving Inflammatory NLRP3 Activation

    29 shares
    Share 12 Tweet 7
  • Radio Tags Alter Flight Kinematics in Neotropical Butterflies Below Standard Weight Limits

    29 shares
    Share 12 Tweet 7
  • Gene Promoter Variations Are Associated with Lobed-Leaf Formation in Siberian Kale

    29 shares
    Share 12 Tweet 7
  • Screening Human Fat Cells Uncovers PLCXD1 Enzyme as Insulin-Regulated Lipid Storage Driver

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Tislelizumab Combined With Chemotherapy Extends Long-Term Survival in Advanced Esophageal Cancer

Global Survey Reveals Marked Variation in Anticoagulation Practices During Adult ECPR

Mutant p53 Reprograms AP-1 Complexes to Drive Cancer Cell Migration

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.