Wednesday, November 18, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News

Vision Transformer Steganography Conceals Twenty Bits Per Image Without Altering Pixels

November 18, 2026
in Science News
Hailey Crawford
By Hailey Crawford Scienmag Editorial Profile - Cybersecurity
Reading Time: 6 mins read
0
Vision Transformer Steganography Conceals Twenty Bits Per Image Without Altering Pixels

Vision Transformer Steganography Conceals Twenty Bits Per Image Without Altering Pixels

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

Researchers Sangeeta Gautam and Manoj Kumar have introduced a coverless information hiding methodology that utilizes a pre-trained Vision Transformer to securely convey confidential binary data across digital images without introducing carrier modifications. Published in the Journal of Transformative Technologies and Sustainable Development, the study demonstrates that natural visual representations can hide secret messages by translating intrinsic semantic features into deterministic binary hash sequences. Rather than embedding information directly into pixel matrices or transforming frequency coefficients, the framework selects unaltered cover images from an existing repository whose computed feature hashes correspond to specific segments of a confidential payload. In empirical evaluations, this approach achieved a hiding capacity of up to twenty bits per image, outperforming established coverless techniques that typically offer between fifteen and sixteen bits. By entirely eliminating carrier file modification, the framework avoids creating statistical visual anomalies that conventional steganalysis tools detect, establishing a reliable pathway for covert digital communication across public and private telecommunication channels.

Digital information security increasingly relies on sophisticated methods to safeguard confidential records, communications, and intellectual property from interception across expanding network infrastructures, industrial controls, and cloud environments. Within the discipline of information hiding, techniques diverge fundamentally into digital watermarking and steganography. Digital watermarking prioritizes durability over total secrecy, placing recognizable identifying markers into media files to authenticate ownership and trace unauthorized digital distribution even when files experience heavy compression or editing. In contrast, steganography prioritizes complete imperceptibility, seeking to conceal the very presence of a confidential message within an innocuous cover medium. Traditional steganographic methods alter host images, introducing subtle numerical shifts across spatial pixel distributions or frequency domain transform coefficients. Although these spatial adjustments may remain imperceptible to human vision, they leave subtle mathematical footprints that modern steganalysis algorithms can identify. Coverless steganography bypasses this fundamental vulnerability entirely by refusing to modify carrier images, relying instead on matching secret data directly with the innate characteristics of unaltered photographs.

Over the past decade, researchers have investigated two distinct branches of coverless information hiding: generative synthesis and selection-based retrieval. Generative approaches utilize deep learning frameworks, such as Wasserstein Generative Adversarial Networks, Deep Convolutional GANs, or Super Resolution GANs, to synthesize artificial stego images directly from secret inputs or latent noise vectors. Although generative models achieve sizable hiding capacities, the synthesized visual outputs frequently suffer from reduced visual fidelity, demand substantial computational training, and exhibit severe vulnerability to standard image processing distortions. Conversely, selection-based coverless techniques examine natural image collections to identify pre-existing files whose innate features naturally align with secret bit sequences. Early selection frameworks relied on low-level statistical indicators, such as sub-image pixel averages, Scale-Invariant Feature Transform points, discrete cosine transforms, or localized convolutional neural network activations. However, these earlier selection methods remained constrained by limited data-hiding capacities, high retrieval complexities, and severe sensitivity to geometric transformations such as image rotation, spatial translation, and peripheral cropping.

To overcome the architectural constraints of local receptive fields in convolutional networks, the authors deployed a pre-trained Vision Transformer, designated as ViT-B/16, to serve as the core feature extraction engine. Traditional convolutional neural networks process images through localized filters that capture immediate pixel neighborhoods, which can leave derived representations susceptible to spatial disruptions and local distortions. The Vision Transformer framework departs from localized convolutions by decomposing an input image into a grid of non-overlapping patches, each measuring sixteen by sixteen pixels. These visual patches are flattened, projected linearly into embedding spaces, and processed through multiple multi-head self-attention layers. This self-attention architecture enables the network to evaluate long-range spatial relationships and model global contextual dependencies across the complete visual scene. By capturing comprehensive structural patterns across whole images, the pre-trained ViT-B/16 model produces highly discriminative, informative feature vectors that remain stable across varying illumination levels, compression profiles, and geometric modifications without requiring task-specific parameter retraining.

The operational workflow of the proposed system begins with systematic feature extraction and robust binary hash sequence generation. Each candidate cover image is resized to a standardized dimension of 224 by 224 pixels and normalized across all color channels to scale pixel intensities between negative one and positive one. The normalized image is passed through the pre-trained Vision Transformer to extract a high-dimensional three-dimensional feature map comprising channel, height, and width coordinates. A linear mapping function then unfolds this three-dimensional tensor into a continuous one-dimensional feature representation vector. This linearized feature vector is subsequently partitioned into twenty equal-sized, non-overlapping blocks. For each individual block, the algorithm computes an arithmetic mean coefficient that summarizes the aggregate feature intensity of that segment. Finally, adjacent mean coefficients are compared sequentially: if the mean of a succeeding block is greater than or equal to the mean of the current block, a binary value of one is appended to the hash sequence; otherwise, a binary zero is recorded, producing a compact hash.

To manage the matching process between binary secret data and large collections of candidate photographs, the authors constructed a high-efficiency inverted index structure. In conventional selection-based steganography, searching a massive image repository through linear scans introduces an algorithmic time complexity of O(n), creating prohibitive computational delays as database volumes expand. The inverted index circumvents linear latency by mapping each unique binary hash sequence directly to a list of corresponding image identification numbers within the dataset. This indexed organization reduces retrieval latency to between O(1) and O(log n), providing near-instantaneous query responses during data transmission preparation. The inverted index also preserves the unique identification numbers of all cataloged images, ensuring that multiple images mapped to identical hash sequences can be selected systematically. Furthermore, retaining explicit identifier ordering within the index prevents transmission sequence scrambling, ensuring that the receiver can reconstruct multi-segment messages in their exact, intended chronological order without ambiguity.

The secret data hiding and extraction procedure operates systematically using the established inverted index and hash generation algorithms. A sender begins with a confidential binary bitstream, which is partitioned into uniform segments matching the predetermined hash length of twenty bits. If the total bit length of the secret message is not evenly divisible by twenty, standard zero-padding is appended to the concluding binary segment to satisfy structural requirements. For each twenty-bit payload segment, the sender queries the inverted index to identify an unaltered carrier image whose computed hash sequence precisely matches the binary segment. The selected images, designated as stego images, are then transmitted sequentially to the intended recipient through ordinary communication channels. Upon receiving the ordered sequence of stego images, the recipient executes the identical ViT-B/16 feature extraction and hash generation algorithm on each received file. Concatenating the reconstructed twenty-bit hashes and trimming the padded zeros from the final segment fully recovers the original secret payload without transmitting auxiliary indexing keys.

Experimental validation of the framework was conducted on a personal computing platform configured with an AMD Ryzen 7 7840HS processor, an NVIDIA GeForce RTX 4060 graphics processing unit with eight gigabytes of dedicated memory, thirty-two gigabytes of system RAM, and a 64-bit Windows 11 operating system, using Python programming environments. The researchers evaluated the model across diverse visual benchmarks, sampling twenty-five nature and wildlife classes from the ImageNet dataset and fifty diverse object categories from the Caltech-256 repository. The selected subsets encompassed complex visual distributions, including marine organisms, avian species, terrestrial animals, manufactured structures, indoor environments, and abstract scenes. Data-hiding capacity tests demonstrated that transmitting a ten-byte message required only four carrier images under the proposed method, compared to five to seven images required by competing frameworks. Transmitting a one-kilobyte secret payload required four hundred images, whereas earlier methods developed by Meng, Li, Zhang, and Liu demanded between 512 and 548 images to transmit equivalent data volumes.

The empirical evaluation thoroughly examined the resilience of the Vision Transformer hash sequences against both non-geometric image processing attacks and geometric distortions. When subjected to non-geometric disruptions, including Gaussian filtering, median filtering, mean filtering, JPEG compression, and scaling operations, the proposed method maintained a retrieval accuracy of 100 percent across both ImageNet and Caltech-256 subsets. Under aggressive noise addition, including Gaussian noise with a variance parameter of 0.005 and speckle noise with a variance of 0.05, the Vision Transformer approach preserved perfect 100 percent extraction accuracy, substantially outperforming a DenseNet-161 baseline which dropped to 58.85 percent under identical Gaussian noise conditions. In geometric evaluations involving center cropping, edge cropping, and horizontal translation, the proposed model sustained near-perfect recovery. Even when carrier images experienced rotational transformations of thirty degrees, the framework achieved extraction accuracies exceeding 80 percent on complex subsets and 100 percent on benchmark comparisons, surpassing competing deep convolutional architectures whose accuracy degraded sharply under severe geometric manipulation.

Despite its demonstrated operational advantages in imperceptibility and attack resilience, the authors outline several practical constraints that govern real-world deployment. The framework exhibits an absolute hiding capacity threshold of twenty bits per image, meaning that large data transfers necessitate transmitting substantial volumes of carrier images, which scales linearly with message size. Additionally, the system remains strictly dependent on the scale and visual diversity of the underlying image database; insufficient visual variety can restrict unique hash generation and elevate the risk of hash collisions. Security also relies on protecting the integrity of the shared image repository, as unauthorized third-party access to the underlying image pool could facilitate reverse-engineering of message selections. Future research directions proposed by the authors include developing advanced multi-scale feature fusion layers, implementing adaptive hashing strategies across varied image domains, and investigating formal legal compliance frameworks for covert communication across highly regulated environments.

Subject of Research: Coverless information hiding using pre-trained Vision Transformers and inverted indexing

Article Title: Feature-Based Coverless Information Hiding Using Vision Transformer

Article References: Gautam, S., & Kumar, M. (2025). Feature-Based Coverless Information Hiding Using Vision Transformer. Journal of Transformative Technologies and Sustainable Development, 9(1), Article 6. https://doi.org/10.1007/s41314-025-00072-7

Image Credits: AI Generated

DOI: 10.1007/s41314-025-00072-7

Keywords: coverless steganography, information hiding, vision transformer, ViT-B/16, inverted index, image hashing, digital steganography, covert communication, ImageNet, Caltech-256, geometric attack robustness, JPEG compression resistance, feature extraction, cybersecurity, data concealment, steganalysis resistance, self-attention mechanism

Cite Scienmag News

Hailey Crawford. (November 18, 2026). Vision Transformer Steganography Conceals Twenty Bits Per Image Without Altering Pixels. Scienmag. https://scienmag.com/vision-transformer-steganography-conceals-twenty-bits-per-image-without-altering-pixels/

Hailey Crawford. "Vision Transformer Steganography Conceals Twenty Bits Per Image Without Altering Pixels." Scienmag, 18 November 2026, https://scienmag.com/vision-transformer-steganography-conceals-twenty-bits-per-image-without-altering-pixels/. Accessed 18 November 2026.

Hailey Crawford. "Vision Transformer Steganography Conceals Twenty Bits Per Image Without Altering Pixels." Scienmag. November 18, 2026. https://scienmag.com/vision-transformer-steganography-conceals-twenty-bits-per-image-without-altering-pixels/

Tags: Caltech-256coverless steganographycovert communicationcybersecuritydata concealmentdigital steganographyfeature extractiongeometric attack robustnessimage hashingImageNetinformation hidinginverted indexJPEG compression resistanceself-attention mechanismsteganalysis resistancevision transformerViT-B/16
Share26Tweet16
Previous Post

Tropical Cyclones and Atmospheric Waves Drive August Heat Waves in Western Japan

Next Post

Dynamic Subarray Partitioning and Clustering Suppress Sidelobes in Linear Radar Arrays

Related Posts

Genomic Researchers Call for Open Collaborative Transposable Element Infrastructure
Science News

Genomic Researchers Call for Open Collaborative Transposable Element Infrastructure

November 18, 2026
Ovary Removal in Aged Mice Spares Knee Cartilage Despite Bone Loss
Biology

Ovary Removal in Aged Mice Spares Knee Cartilage Despite Bone Loss

November 18, 2026
New GPU-Powered Solver Enables Real-Time Industrial 3D Physics Simulations
Mathematics

New GPU-Powered Solver Enables Real-Time Industrial 3D Physics Simulations

November 18, 2026
Seed-to-Seed Techniques and Vernalization Maximize Tropical Botanical Shallot Seed Yields
Agriculture

Seed-to-Seed Techniques and Vernalization Maximize Tropical Botanical Shallot Seed Yields

November 18, 2026
Stroke Recovery: Upper Limb Capacity and Real-World Usage Diverge Over Time
Medicine

Stroke Recovery: Upper Limb Capacity and Real-World Usage Diverge Over Time

November 18, 2026
Template-Guided Patient-Specific Implants Reduce Chronic Temporomandibular Joint Dislocation Symptoms
Medicine

Template-Guided Patient-Specific Implants Reduce Chronic Temporomandibular Joint Dislocation Symptoms

November 18, 2026
Next Post
Dynamic Subarray Partitioning and Clustering Suppress Sidelobes in Linear Radar Arrays

Dynamic Subarray Partitioning and Clustering Suppress Sidelobes in Linear Radar Arrays

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Genomic Researchers Call for Open Collaborative Transposable Element Infrastructure
  • Ovary Removal in Aged Mice Spares Knee Cartilage Despite Bone Loss
  • New GPU-Powered Solver Enables Real-Time Industrial 3D Physics Simulations
  • Seed-to-Seed Techniques and Vernalization Maximize Tropical Botanical Shallot Seed Yields

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

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

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

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
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading