Page 688

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

AI-Based and Compressed-Domain Video Steganography: A

Systematic Review and Comparative Analysis (2017–2025)

Anamika Saini

, Kavita Rathi

Department of Computer Science and Engineering, DeenBandhu ChhotuRam University of Science &

Technology, Sonipat, Haryana, India

DOI:

https://doi.org/10.51583/IJLTEMAS.2026.150600054

Received: 14 June 2026; Accepted: 19 June 2026; Published: 04 July 2026

ABSTRACT

Artificial intelligence has significantly transformed video steganography by improving the security,

robustness, and adaptability of secret data embedding techniques. This paper presents a systematic review

and comparative analysis of AI-based and compressed-domain video steganography techniques

published between 2017 and 2025. Relevant studies were collected from major scientific databases and

critically analyzed based on embedding domain, deep learning architecture, dataset, payload capacity,

peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), mean square error (MSE),

computational complexity, robustness, and resistance to steganalysis attacks. The review highlights the

advantages and limitations of convolutional neural network (CNN), recurrent neural network (RNN),

generative adversarial network (GAN), transformer-based, and compressed-domain approaches.

Comparative analysis indicates that AI-driven techniques generally achieve superior imperceptibility and

security compared to conventional methods, although higher computational requirements remain a

significant challenge. The study further identifies existing research gaps, discusses practical

implementation challenges, and outlines future research directions including lightweight deep learning

models, explainable AI, federated learning, and real-time secure multimedia communication systems.

Keywords: Deep Learning Steganography, GAN, CNN, Video Steganography, Compressed Domain

Embedding, Diffusion Models, Multimedia Security

INTRODUCTION

The rapid expansion of cloud computing, intelligent multimedia applications, and digital communication

networks has significantly increased the demand for secure information transmission techniques. Among

various information security approaches, steganography has emerged as an effective method for concealing

confidential data within digital media while maintaining the visual quality of the cover content. Traditional

steganographic techniques, such as Least Significant Bit (LSB) substitution and Pixel Value Differencing

(PVD), offer high embedding capacity and implementation simplicity. However, these methods often suffer

from limited robustness and are vulnerable to modern statistical and machine learning-based steganalysis

attacks [23], [24].

Recent advancements in Artificial Intelligence (AI) and deep learning have revolutionized the field of

steganography by enabling adaptive and data-driven embedding strategies. Deep neural network

architectures, including Convolutional Neural Networks (CNNs), Generative Adversarial Networks (GANs),

Vision Transformers (ViTs), and Diffusion Models, have demonstrated remarkable capabilities in

automatically learning optimal embedding regions while minimizing visual distortion [1]-[5]. These

intelligent approaches significantly enhance imperceptibility, payload capacity, robustness, and resistance

against sophisticated steganalysis techniques, making them suitable for next-generation multimedia security

applications.

In parallel, compressed-domain video steganography has attracted considerable attention due to its ability to

embed secret information directly into compressed video streams. By exploiting video-specific features such

Page 689

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

as motion vectors, Discrete Cosine Transform (DCT) coefficients, prediction residuals, and temporal

redundancies, compressed-domain methods achieve improved robustness against recompression, format

conversion, and transmission-related distortions. These characteristics make them highly applicable in

modern cloud-based multimedia communication systems and real-time video-sharing platforms [12], [14],

[18].

This review presents a comprehensive analysis of contemporary AI-driven video steganography and

compressed-domain video steganography techniques. It systematically examines recent developments,

underlying methodologies, performance metrics, and security challenges associated with intelligent

multimedia hiding systems. Furthermore, a comparative evaluation of state-of-the-art approaches is provided

to highlight their strengths, limitations, and future research directions in secure multimedia communication

[7], [15], [20], [21].

Unlike conventional research articles that propose a new embedding algorithm, this work presents a

systematic review and comparative analysis of recent AI-based and compressed-domain video

steganography techniques published between 2017 and 2025. The objective of this study is to critically

evaluate existing approaches using common performance metrics, identify their strengths and limitations,

analyze computational and security aspects, and highlight potential research opportunities for future

developments in secure multimedia communication.

Objective of This Review

1. To analyze AI-based video steganography techniques.

2. To compare CNN, GAN, transformer, and diffusion-based methods.

3. To evaluate compressed-domain embedding approaches.

4. To identify research gaps in intelligent multimedia security.

5. To discuss future research opportunities in secure video communication.

RESEARCH METHODOLOGY

This review adopts a systematic literature review (SLR) approach to analyze recent advancements in AI-

based video steganography and compressed-domain video steganography. The objective is to identify,

evaluate, and compare state-of-the-art techniques proposed for secure multimedia communication. The

review process involved defining search strategies, selecting relevant studies, and applying predefined

inclusion and exclusion criteria to ensure the quality and relevance of the collected literature. This review

follows a structured methodology for identifying, screening, and analyzing relevant research articles.

Publications between 2017 and 2025 were collected from leading digital libraries including IEEE Xplore,

ScienceDirect, SpringerLink, ACM Digital Library, and Scopus-indexed journals. Keywords such as "Video

Steganography", "Deep Learning", "CNN", "GAN", "Compressed Domain", "H.264", "HEVC", and

"Artificial Intelligence" were used during the literature search. Studies were selected based on their relevance

to AI-enabled video steganography, availability of experimental evaluation, and publication quality.

Duplicate records, non-English articles, and studies unrelated to video steganography were excluded. The

selected papers were comparatively analyzed based on embedding strategy, datasets, payload capacity,

PSNR, SSIM, MSE, robustness, computational complexity, and resistance to steganalysis.

Data Sources

Relevant studies were collected from leading scientific databases and digital libraries to ensure

comprehensive coverage of high-quality research publications. The primary sources included IEEE Xplore,

SpringerLink, Elsevier ScienceDirect, ACM Digital Library, MDPI, and Google Scholar. These databases

were selected because they provide access to peer-reviewed journals, conference proceedings, and recent

Page 690

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

developments in multimedia security, artificial intelligence, and information hiding [7], [12].

Search Keywords

A structured keyword-based search strategy was employed to identify studies related to intelligent

steganography and secure multimedia communication. The search queries included terms such as “Deep

Learning Steganography,” “GAN-based Image Hiding,” “CNN Video Steganography,” “Transformer

Steganography,” “Diffusion Model Image Hiding,” “Motion Vector Embedding,” and “Compressed-Domain

Steganography.” Various combinations of these keywords were used to maximize the retrieval of relevant

publications and minimize the risk of missing significant contributions.

Inclusion Criteria

To ensure the relevance and quality of the reviewed literature, studies were selected based on the following

inclusion criteria:

 Publications published between 2017 and 2025.

 Research focusing on AI-driven steganography techniques, including deep learning, GANs,

transformers, and diffusion-based models.

 Peer-reviewed journal articles and conference papers.

 Studies providing experimental evaluations, performance metrics, datasets, or comparative analyses.

 Research addressing image or video steganography for multimedia security applications.

Exclusion Criteria

Studies were excluded from the review if they met any of the following conditions:

 Duplicate publications retrieved from multiple databases.

 Non-English language publications.

 Non-peer-reviewed articles, editorials, blogs, theses, or technical reports.

 Studies focusing exclusively on conventional steganographic methods without incorporating AI or

intelligent optimization techniques.

 Publications lacking sufficient experimental validation, performance evaluation, or methodological

details.

The selected studies were subsequently analyzed and compared based on embedding capacity,

imperceptibility, and robustness, security against steganalysis, computational complexity, datasets used, and

application domains. This systematic methodology ensures a comprehensive and unbiased assessment of

recent advancements in intelligent multimedia steganography.

LITERATURE REVIEW OF EXISTING STUDIES

A comprehensive review of existing literature was conducted to examine the recent developments in AI-

based and compressed-domain steganography techniques. The selected studies were obtained from reputed

databases, including IEEE Xplore, SpringerLink, Elsevier ScienceDirect, ACM Digital Library, and Google

Scholar. The reviewed papers primarily focus on deep learning architectures such as Convolutional Neural

Networks (CNNs), Generative Adversarial Networks (GANs), transformers, diffusion models, and advanced

video steganography techniques.

Page 691

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

The objective of this literature review is to identify the major contributions, strengths, limitations, and

research trends in intelligent multimedia security. Table 1 presents a summary of representative studies

published between 2017 and 2025, highlighting their methodologies and key findings. Table 1 summarizes

the major AI-based and compressed-domain steganography techniques reported between 2017 and 2025.

The selected studies highlight recent advancements in deep learning, GAN-based hiding systems,

transformer architectures, and secure video steganography frameworks [1]– [21].

Table 1. Summary of AI-Based and Compressed-Domain Steganography Studies (2017–2025)

Sr.

No.

Authors

Year

Title / Technique

Key Contribution

Source

Shumeet

Baluja

2017

Hiding Images in Plain

Sight: Deep Steganography

First deep neural

network framework

for image-in-image

hiding

NeurIPS

Jamie Hayes,

George Danezis

2017

Generating Steganographic

Images via Adversarial

Training

Alice-Bob-Eve

adversarial

steganography

framework

NeurIPS

Denis

Volkhonskiy et

al.

2017

Generative Adversarial

Networks for Video

steganography

Early GAN-based

steganography model

OpenReview

Zhu, Kaplan,

Johnson & Fei-

Fei

2018

HiDDeN: Hiding Data With

Deep Networks

End-to-end deep

learning framework

for data hiding

ECCV

Ru Zhang,

Shiqi Dong,

Jianyi Liu

2019

Invisible Steganography via

GANs

GAN-based secure

image hiding with

improved invisibility

Multimedia

Tools &

Applications

Kevin A.

Zhang et al.

2019

SteganoGAN: High-

Capacity Video

steganography with GANs

High payload

capacity and

imperceptibility

arXiv / MIT

DAI Lab

Duan et al.

2019

Reversible Video

steganography Based on U-

Net Structure

U-Net based

reversible image

hiding

IEEE Access

Boroumand et

al.

2019

Deep Residual Network for

Steganalysis

Improved detection

of hidden information

IEEE TIFS

Rehman et al.

2019

End-to-End Trained CNN

Encoder–Decoder Networks

Deep encoder-

decoder video

steganography

ECCV

Workshop

Tabares-Soto

et al.

2019

Deep Learning Applied to

Steganalysis of Digital

Images

Systematic review of

deep steganalysis

methods

IEEE Access

Baluja

2020

Hiding Images Within

Images

Enhanced deep

steganography

architecture

IEEE

TPAMI

Page 692

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

The reviewed studies indicate a clear transition from traditional embedding methods toward intelligent data-

driven approaches. Recent research has focused on improving imperceptibility, payload capacity, robustness,

and resistance against steganalysis using deep neural networks and compressed-domain video embedding

strategies [1]–[21].

Furthermore, the literature indicates that AI-driven steganography methods generally achieve superior

performance compared to traditional approaches in terms of security, feature learning, and resistance to

steganalysis. However, challenges such as high computational complexity, large training data requirements,

and real-time implementation constraints remain significant research concerns. These observations motivate

the need for further investigation into efficient, robust, and scalable steganographic solutions for next-

generation multimedia communication systems.

Although AI-based video steganography techniques provide superior security and imperceptibility, they

often require significant computational resources due to deep neural network training and inference. CNN-,

GAN-, and transformer-based models generally demand high-performance GPUs, large annotated datasets,

and increased execution time compared to traditional compressed-domain approaches. These computational

requirements limit their deployment in real-time multimedia applications, mobile devices, and edge

computing environments. Future research should therefore focus on lightweight architectures capable of

achieving an optimal balance between security, computational efficiency, and embedding performance

Modern AI-based video steganography techniques are designed to improve resistance against statistical and

deep learning-based steganalysis attacks. Adaptive embedding strategies and feature learning mechanisms

reduce detectable artifacts within stego videos, thereby improving communication security. However, recent

CNN-based steganalysis models have demonstrated increasing detection capability, indicating that

developing more robust and adaptive embedding frameworks remains an important research challenge.

TAXONOMY OF AI-BASED VIDEO STEGANOGRAPHY

CNN-Based Steganography

Convolutional Neural Networks (CNNs) have become one of the most widely adopted deep learning

architectures in steganography. Unlike conventional embedding methods that rely on manually designed

rules, CNN-based approaches automatically learn discriminative features from training data and identify

Chen, Wang et

al.

2020

High-Capacity Robust

Video steganography via

Adversarial Network

Robust adversarial

image hiding

KSII

Transactions

Chen, Xing

& Liu

2020

Technology of Hiding and

Protecting Secret Image

Based on Two-Channel

Deep Hiding Network

Two-channel deep

hiding network

IEEE Access

Zhou et al.

2020

Security Enhancement of

Steganography via

Generative Adversarial

Image

Improved security

using GAN-

generated images

IEEE Signal

Processing

Letters

Li et al.

2021

Secure Steganography for

Hiding Images via GAN

Enhanced security

and image quality

Journal of

Image and

Video

Processing

Wani et al.

2022/2023

Deep Learning Based

Video steganography: A

Review

Comprehensive

review of DL

steganography

WIREs

Data

Mining

Page 693

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

suitable regions for secret data embedding. By exploiting spatial characteristics within images and video

frames, CNN models can achieve a favorable balance between embedding capacity and visual quality [8],

[10].

A major advantage of CNN-based steganography is its ability to adaptively determine embedding locations

based on content characteristics, resulting in improved imperceptibility and enhanced resistance to

steganalysis attacks. Furthermore, CNNs are effective at extracting complex features that are difficult to

capture using traditional methods [28], [29]. However, these models generally require large annotated

datasets for effective training and often depend on powerful GPU resources. The training process can also

be computationally intensive and time-consuming, particularly for large-scale multimedia datasets [1], [4],

[6].

GAN-Based Steganography

Generative Adversarial Networks (GANs) have introduced a new paradigm in steganography by employing

a generator-discriminator framework. In this approach, the generator creates stego-images that conceal secret

information, while the discriminator attempts to distinguish between original and modified content. Through

this adversarial learning process, GANs continuously improve the realism and security of generated stego-

media.

GAN-based methods are known for producing highly imperceptible stego-images that closely resemble the

original content. Their ability to learn adaptive embedding strategies makes them more resistant to detection

by modern steganalysis techniques [27]. Despite these advantages, GANs are often difficult to train due to

issues such as mode collapse and training instability. In addition, they require substantial computational

resources, large training datasets, and careful parameter tuning to achieve optimal performance [2], [3], [5],

[13].

Transformer-Based Steganography

Transformer architectures have recently gained attention in multimedia steganography because of their

capability to model long-range dependencies and global contextual information. Unlike CNNs, which

primarily focus on local spatial features, transformers utilize self-attention mechanisms to analyze

relationships across the entire image or video frame [18], [20].

The ability to capture global context enables transformer-based methods to identify more suitable embedding

regions and improve the overall robustness of hidden communication. These models often demonstrate

superior feature representation and stronger resistance to various attacks. Nevertheless, transformer

architectures typically involve a large number of parameters, resulting in high computational costs and

significant memory consumption. Their training also requires extensive datasets and powerful hardware

platforms, which may limit practical deployment in resource-constrained environments [15], [18], [20].

Diffusion-Based Steganography

Diffusion models represent one of the latest advancements in AI-driven steganography. These models

generate stego-images through a gradual denoising process that reconstructs high-quality visual content

while preserving hidden information [21], [22]. The iterative nature of diffusion models allows them to create

highly realistic images with minimal visual artifacts.

One of the primary strengths of diffusion-based steganography is its exceptional visual quality and strong

resistance to detection. The generated stego-images often exhibit a high degree of naturalness, making secret

communication more secure against advanced steganalysis methods. However, these benefits come at the

cost of increased computational complexity. Diffusion models generally require lengthy training and

inference times, as well as high-performance hardware resources, which can pose challenges for real-time

applications.

Page 694

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

Compressed-Domain Video Steganography

Compressed-domain video steganography has emerged as an efficient approach for secure multimedia

communication because it embeds secret information directly into compressed video streams rather than

modifying raw video frames. Unlike spatial-domain techniques, which alter pixel values before compression,

compressed-domain methods exploit the structural components generated during video encoding. This

approach not only reduces computational overhead but also improves compatibility with modern video

transmission systems.

Contemporary video compression standards such as H.264/AVC, H.265/HEVC, and AV1 provide multiple

embedding opportunities within the encoded bitstream. Researchers utilize these components to conceal

secret information while maintaining acceptable video quality and minimizing the risk of detection. Since

the embedding process occurs after or during compression, compressed-domain steganography is generally

more robust against recompression and format conversion attacks. Several components of compressed video

streams are commonly used for data hiding:

Motion Vectors

Motion vectors represent the movement of objects between consecutive video frames. By slightly modifying

selected motion vector values, secret information can be embedded without causing noticeable visual

distortion. Motion vector-based methods are widely used because they exploit temporal redundancy and offer

good embedding capacity.

DCT Coefficients

Discrete Cosine Transform (DCT) coefficients are generated during video compression to represent

frequency-domain information. Secret bits can be embedded by adjusting selected coefficients, particularly

those located in middle-frequency regions where visual changes are less perceptible. DCT-based techniques

provide a balance between imperceptibility and robustness.

Quantization Parameters (QPs)

Quantization is a critical stage of video compression that controls the trade-off between video quality and

compression efficiency. Small modifications in quantization parameters can be used to encode secret

information while preserving the overall visual quality of the video.

Prediction Modes

Modern video codecs employ intra-frame and inter-frame prediction techniques to reduce redundancy.

Certain prediction mode selections can be manipulated to carry hidden information without significantly

affecting coding performance, making them suitable for covert communication applications.

Advantages of Compressed-Domain Video Steganography

Improved Resistance to Recompression: Since the secret data is embedded within the compressed structure

itself, these techniques are generally more resilient to recompression and transcoding operations commonly

performed by video-sharing platforms.

Suitability for Real-Time Applications: Embedding information directly into compressed streams reduces

processing requirements, making these methods suitable for live video transmission, video conferencing, and

cloud-based multimedia services.

Lower Storage and Bandwidth Overhead: Because the embedding process does not require complete

decompression and reconstruction of video frames, storage requirements and transmission overhead are

minimized.

Page 695

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

Enhanced Security: The hidden information is distributed within codec-specific parameters, making

detection more difficult for attackers who rely on conventional steganalysis techniques.

Limitations of Compressed-Domain Video Steganography

Codec Dependency: Most embedding techniques are designed for specific video compression standards. A

method developed for H.264 may not function effectively with H.265 or AV1, limiting interoperability

across different platforms.

Synchronization Complexity: Accurate extraction of hidden information often depends on maintaining

synchronization between encoder and decoder operations. Any alteration in the video stream may lead to

data recovery errors.

Implementation Difficulty: Understanding and modifying codec internals requires specialized expertise in

video compression algorithms, making implementation significantly more complex than traditional image-

based steganography techniques.

Limited Embedding Flexibility: To preserve compression efficiency and visual quality, only certain codec

parameters can be modified, which may restrict embedding capacity in some applications.

Overall, compressed-domain video steganography represents a promising solution for secure multimedia

communication due to its robustness, efficiency, and compatibility with modern video coding standards.

However, challenges related to codec dependency, implementation complexity, and synchronization must be

addressed to further enhance its practical deployment in real-world communication systems [12], [14], [18].

Dataset Analysis and Experimental Overview

The performance and reliability of steganographic systems largely depend on the quality and diversity of the

datasets used during training and evaluation. In recent years, researchers have increasingly relied on large-

scale image and video datasets to develop intelligent steganography models capable of learning complex

spatial and temporal patterns. These datasets provide a variety of visual characteristics, including textures,

object distributions, motion information, and scene variations, which are essential for training robust deep

learning models. For video steganography, datasets such as ImageNet, BOSSBase, COCO, and CelebA are

frequently employed due to their extensive visual diversity and availability of high-quality images. These

datasets enable researchers to evaluate embedding capacity, visual imperceptibility, and resistance against

steganalysis attacks under different conditions.

Similarly, video steganography research utilizes datasets such as HMDB51, UCF101, and Vimeo-90K.

These datasets contain dynamic scenes and motion-rich content that help evaluate temporal embedding

strategies and robustness against video compression and transmission distortions. In specialized domains

such as healthcare, Medical MRI datasets are increasingly used to investigate secure transmission of sensitive

medical information while preserving diagnostic image quality. Table 1 presents some of the most commonly

used datasets in contemporary AI-driven video steganography research. Various benchmark datasets are used

for training and evaluating intelligent steganography systems. These datasets support image hiding, video

steganography, steganalysis, and multimedia security applications [7], [15], [18].

Table 2. Commonly Used Datasets in AI-Based Video steganography Research

Sr.

No.

Dataset

Approximate Files

Usage

ImageNet

Millions of Images

Training deep learning and feature extraction models

BOSSBase

10,000

Images

Benchmark dataset for steganography and steganalysis

evaluation

Page 696

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

The selection of appropriate datasets plays a critical role in determining model performance, generalization

capability, and robustness against attacks [7], [15]. These datasets collectively support the development and

evaluation of modern steganographic systems by providing diverse multimedia content. Their widespread

adoption has contributed significantly to advancements in deep learning-based information hiding, enabling

researchers to design more secure, robust, and imperceptible steganography techniques for real-world

applications.

Performance Evaluation Metrics

The effectiveness of steganographic techniques is commonly assessed using quantitative performance

metrics that measure visual quality, distortion, and embedding capability. In both video steganography,

metrics such as Peak Signal-to-Noise Ratio (PSNR), Mean Squared Error (MSE), Root Mean Squared Error

(RMSE), and embedding capacity are widely used to evaluate the quality of stego-media and the efficiency

of information hiding algorithms.

Peak Signal-to-Noise Ratio (PSNR)

One of the most important quality assessment metrics. Higher PSNR values indicate that the stego-image or

stego-video closely resembles the original cover media, resulting in better imperceptibility. In general, PSNR

values above 40 dB are considered acceptable for secure multimedia communication, while values exceeding

50 dB indicate excellent visual quality.

Mean Squared Error (MSE)

It measures the average squared difference between the original and stego content. Lower MSE values

indicate less distortion and better preservation of visual information after data embedding.

Root Mean Squared Error (RMSE)

It is derived from MSE and provides a more interpretable measure of distortion by expressing the error in

the same unit as the original pixel values. Lower RMSE values correspond to higher-quality stego content.

Embedding Capacity

It refers to the amount of secret information that can be hidden within a cover image or video while

maintaining acceptable visual quality and security. A higher embedding capacity is desirable; however,

increasing payload size may affect imperceptibility and robustness.

COCO

Dataset

330K

Images

GAN-based video steganography and object-rich image

analysis

CelebA

200K

Images

Facial video steganography and identity-preserving

data hiding

HMDB51

6,700

Videos

Motion-based video steganography and temporal

feature analysis

UCF101

13,000

Videos

Human action videos for video hiding and classification

tasks

Vimeo-

90K

90,000

Video

Sequences

Temporal consistency learning and video embedding

research

Medical MRI

Dataset

5,000

Images

Secure healthcare communication and medical data

protection

Page 697

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

Table 3 presents a comparative summary of performance metrics reported for various AI-based and

compressed-domain steganography techniques. Performance of steganographic systems is commonly

evaluated using objective quality and distortion metrics such as PSNR, MSE, RMSE, and embedding

capacity [5], [9], [13], [18].

Table 3. Performance Evaluation Metrics of AI-Based and Compressed-Domain Steganography

Techniques

Higher PSNR and lower MSE/RMSE values generally indicate improved visual quality and imperceptibility

of the stego content [5], [13], [18]. The comparative analysis indicates that AI-driven approaches generally

achieve superior visual quality and lower distortion compared to conventional techniques. Among the

reviewed methods, diffusion-based steganography demonstrates the highest average PSNR and the lowest

error values, reflecting its ability to generate highly realistic stego-images. GAN-based and transformer-

based approaches also exhibit strong performance due to their advanced feature-learning capabilities and

adaptive embedding mechanisms. In contrast, compressed-domain video steganography techniques provide

moderate visual quality but offer additional advantages such as improved robustness against recompression

and suitability for real-time multimedia transmission.

Overall, the results suggest that modern AI-based steganography methods have significantly improved the

balance between imperceptibility, embedding capacity, and security, making them promising candidates for

next-generation secure multimedia communication systems.

Comparative Analysis of Video Steganography Techniques

The rapid evolution of artificial intelligence has significantly transformed modern steganography by

introducing intelligent embedding mechanisms capable of improving security, imperceptibility, and

resistance against detection. At the same time, compressed-domain video steganography continues to play

an important role in multimedia communication due to its efficiency and compatibility with contemporary

video coding standards. A comparative analysis of these approaches helps identify their strengths,

limitations, and suitability for different application scenarios.

AI-based steganographic techniques generally outperform conventional methods in terms of security and

adaptability. Deep learning models are capable of automatically learning optimal embedding locations from

large datasets, thereby reducing the likelihood of detection by steganalysis systems. However, these

Sr.

No.

Technique

Average

PSNR

Average

MSE

Average

RMSE

Capacity

CNN-Based

Stego

42–56

0.002–0.009

0.04–0.09

Medium

GAN-Based

Stego

50–65

0.001–0.004

0.02–0.06

High

Transformer

Stego

48–63

0.001–0.005

0.02–0.07

High

Diffusion Stego

55–70

0.0005–

0.003

0.01–0.05

Medium

Motion Vector

Stego

40–54

0.003–0.010

0.05–0.10

Medium

Compressed

DCT Stego

43–57

0.002–0.008

0.04–0.09

Medium

Page 698

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

improvements often come at the cost of increased computational complexity and training requirements.

CNN-based steganography provides a good balance between security and computational efficiency. By

learning spatial features from multimedia content, CNN models achieve high security and reasonable

robustness while maintaining moderate suitability for practical deployment. Nevertheless, their performance

depends heavily on the availability of large training datasets and computational resources.

GAN-based steganography offers enhanced security and imperceptibility through adversarial learning. The

interaction between generator and discriminator networks enables the creation of highly realistic stego-

images that are difficult to distinguish from original content. As a result, GAN-based methods demonstrate

superior resistance against steganalysis attacks. However, their training process is computationally expensive

and often suffers from stability challenges.

Transformer-based approaches further improve feature representation by capturing global contextual

relationships across the entire image or video frame. This capability enhances robustness and embedding

effectiveness, particularly in complex multimedia environments. Despite these advantages, transformer

architectures require substantial computational resources, large-scale datasets, and extensive training time,

which limits their applicability in real-time systems.

Diffusion-based steganography represents one of the most advanced developments in AI-driven information

hiding. These models generate highly realistic stego-images through iterative denoising processes, resulting

in exceptional visual quality and strong resistance to detection. However, their computational demands are

significantly higher than those of CNNs, GANs, and transformers, making real-time implementation

challenging.

In contrast, compressed-domain video steganography focuses on embedding information directly into

encoded video streams. These methods are generally more efficient and suitable for practical communication

systems because they avoid the need for extensive model training.

Motion vector embedding techniques exploit temporal information within compressed video streams and

offer good robustness against recompression and transmission-related distortions. Their relatively low

computational requirements make them suitable for real-time video applications, although their security level

is generally lower than that of AI-based approaches.

Similarly, DCT-based embedding methods utilize frequency-domain coefficients generated during video

compression. These techniques provide a reasonable balance between robustness, visual quality, and

implementation complexity. Due to their efficiency and compatibility with existing video codecs, DCT-

based methods remain widely used in practical multimedia security systems.

Table 4 compares AI-based and compressed-domain steganography techniques based on security,

robustness, computational complexity, training requirements, and real-time applicability [1]-[21].

Table 4. Comparison of AI-Based Steganography Techniques

Sr.

No.

Method

Security

Robustness

Complexity

Training Cost

Real-Time

Suitability

CNN

High

Medium

High

Moderate

GAN

Very High

High

Very High

Moderate

Transformer

High

Extremely

High

Very High

Low

Diffusion

Models

Extremely

High

Extremely

High

Extremely High

Low

Page 699

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

GAN and diffusion-based approaches provide superior security and visual quality, whereas compressed-

domain techniques offer better real-time performance and lower computational overhead [12], [14], [18],

[21]. The comparison reveals that AI-based techniques generally provide superior security and

imperceptibility, making them suitable for applications requiring advanced protection against steganalysis.

Among these methods, diffusion models and GANs demonstrate the strongest security performance but incur

substantial computational and training costs. On the other hand, compressed-domain approaches offer lower

implementation costs and excellent real-time performance, making them more practical for video streaming,

cloud communication, and resource-constrained environments.

Therefore, the selection of an appropriate steganographic technique depends on the specific application

requirements. AI-driven methods are preferable when security and robustness are the primary objectives,

whereas compressed-domain approaches remain attractive for real-time multimedia communication systems

where computational efficiency is a critical consideration.

RESULT AND DISCUSSION

The reviewed studies demonstrate that deep learning has significantly transformed the landscape of video

steganography. CNN-based approaches introduced adaptive feature learning mechanisms that improved

embedding efficiency and enhanced resistance against conventional steganalysis methods. Similarly, GAN-

based techniques achieved superior imperceptibility and visual quality by learning optimal embedding

distributions through adversarial training [1], [2], [5], [13].

Recent advancements in transformer and diffusion-based architectures have further improved the

performance of intelligent steganographic systems. Transformer models effectively capture global contextual

relationships, enabling more robust and secure embedding strategies. Diffusion-based approaches have

shown exceptional visual realism and often achieve higher PSNR values with lower distortion compared to

earlier deep learning methods. However, these benefits come at the expense of increased computational

complexity, extensive training requirements, and high hardware dependency.

Compressed-domain video steganography techniques, particularly those based on motion vectors and DCT

coefficients, have demonstrated strong resilience against recompression and transmission-related distortions.

Their compatibility with modern video coding standards makes them suitable for real-time multimedia

communication environments. Furthermore, hybrid approaches that combine artificial intelligence with

compressed-domain embedding have shown promising improvements in robustness, security, and

embedding efficiency.

Overall, the comparative analysis indicates that AI-driven steganography techniques generally outperform

traditional methods in terms of security, imperceptibility, and adaptability. Nevertheless, challenges related

to computational cost, dataset dependency, model explainability, and real-time deployment remain important

areas for future research [15], [16] [18].

Research Gaps

Despite the remarkable progress achieved in AI-based and compressed-domain steganography, several

research challenges remain unresolved. Most deep learning models operate as black-box systems, making it

difficult to interpret embedding decisions and assess security behavior. In addition, the high computational

requirements of modern architectures limit their deployment in resource-constrained environments. Existing

approaches are heavily dependent on large-scale datasets and powerful GPU infrastructures for effective

training.

Motion Vector

Embedding

Medium

High

Medium

Low

High

DCT-Based

Embedding

Medium

Low

High

Page 700

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

Another significant challenge is the lack of lightweight steganographic frameworks suitable for edge devices,

Internet of Things (IoT) applications, and mobile platforms. Furthermore, limited research has focused on

adaptive compressed-domain embedding techniques capable of maintaining robustness under varying

compression standards and transmission conditions. Real-time secure multimedia communication also

remains an open research problem, as many existing methods fail to balance security, embedding capacity,

and computational efficiency simultaneously. Table 5 summarizes the major research gaps identified from

the reviewed literature. Despite significant advancements, several challenges remain unresolved in intelligent

multimedia steganography systems [7], [15], [17] [18], [21].

Table 5. Major Research Gaps Identified in AI-Based and Compressed-Domain Steganography

Future Scope

The rapid advancement of artificial intelligence and multimedia communication technologies is expected to

create new opportunities for steganography research. Future developments will likely focus on improving

security, computational efficiency, robustness, and real-time applicability of intelligent information hiding

systems. Researchers are increasingly exploring advanced deep learning architectures and adaptive

embedding mechanisms to address the limitations of existing approaches [15], [18], [21], [22]. Several

promising research directions have been identified from the reviewed literature:

 Development of lightweight transformer architectures for resource-constrained environments.

 Real-time GAN-based video steganography for live multimedia communication.

 Compression-aware neural embedding frameworks capable of adapting to different video codecs.

 AI-driven medical video steganography for secure healthcare data transmission.

 Edge-device compatible steganographic systems for IoT and mobile applications.

 Explainable and interpretable steganography frameworks to improve transparency and trustworthiness.

 Quantum-resistant multimedia security models for future secure communication networks.

 Hybrid architectures integrating CNNs, GANs, transformers, and diffusion models.

 Robust steganography techniques resilient to advanced AI-based steganalysis attacks.

 Privacy-preserving multimedia communication systems for cloud and distributed environments.

Future research should focus on developing lightweight deep learning architectures that reduce

computational complexity while maintaining high embedding capacity and imperceptibility. Emerging

technologies such as transformer-based networks, federated learning, explainable artificial intelligence,

blockchain-assisted secure communication, and quantum-inspired steganography have significant potential

to enhance future video steganography systems. Furthermore, improving robustness against advanced AI-

Sr. No.

Research Gap

Description

Lack of Explainability

Most deep models behave as black-box

systems.

High Computational Cost

Training requires GPUs and high

memory resources.

Limited Lightweight Models

Existing models are unsuitable for edge

devices.

Dataset Dependency

Large datasets are mandatory for

training.

Compression Robustness

Limited work on adaptive compressed-

domain hiding.

Page 701

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

driven steganalysis and enabling real-time deployment on resource-constrained devices remain important

research priorities.

Concluding Remark: AI-driven and compressed-domain steganography have demonstrated significant

potential for secure multimedia communication. Future advancements in lightweight deep learning models,

explainable AI, quantum-resistant security, and real-time embedding techniques are expected to further

enhance the effectiveness, robustness, and practical applicability of intelligent steganographic systems [25].

CONCLUSION

This paper presented a systematic review of AI-based and compressed-domain video steganography

techniques published between 2017 and 2025. The reviewed studies demonstrate that artificial intelligence

has significantly transformed modern steganography by enabling adaptive, intelligent, and highly secure data

embedding mechanisms. Deep learning architectures, including CNNs, GANs, transformers, and diffusion

models, have achieved substantial improvements in imperceptibility, robustness, embedding capacity, and

resistance against steganalysis when compared with traditional steganographic approaches [26].

In addition, compressed-domain video steganography techniques utilizing motion vectors, DCT coefficients,

and other codec-specific features have shown strong resilience against recompression and video processing

operations, making them suitable for real-time multimedia communication systems. The comparative

analysis further revealed that while AI-driven methods offer superior security and visual quality, they often

require significant computational resources, extensive training datasets, and complex optimization

procedures.

Despite the remarkable progress achieved in recent years, several challenges remain, including high

computational complexity, limited model explainability, dataset dependency, and constraints related to real-

time deployment. Future research should focus on developing lightweight, explainable, and computationally

efficient steganographic frameworks that can operate effectively in cloud, edge, and resource-constrained

environments. The integration of advanced artificial intelligence techniques with robust compressed-domain

embedding strategies is expected to play a vital role in the development of next-generation secure multimedia

communication systems.

Overall, this systematic review demonstrates that AI-based video steganography has achieved significant

progress in enhancing secure multimedia communication. Despite improvements in embedding quality,

robustness, and adaptive learning, challenges related to computational efficiency, scalability, dataset

diversity, and resistance to advanced steganalysis remain open research problems. Addressing these issues

will facilitate the development of practical, secure, and intelligent video steganography systems suitable for

future real-world applications.

REFERENCES

1. Baluja, S. (2017). Hiding images in plain sight: Deep steganography. In Advances in Neural

Information Processing Systems, 30, 2069–2079.

https://scholar.google.com/scholar?q=Hiding+Images+in+Plain+Sight+Deep+Steganography

2. Hayes, J., & Danezis, G. (2017). Generating steganographic images via adversarial training. In

Advances in Neural Information Processing Systems, 30, 1954–1963.

https://scholar.google.com/scholar?q=Generating+Steganographic+Images+via+Adversarial+Trainin

3. Volkhonskiy, D., Nazarov, I., Burnaev, E., & Borisenko, A. (2017). Steganographic generative

adversarial networks. arXiv preprint arXiv:1703.05502.

https://scholar.google.com/scholar?q=Steganographic+Generative+Adversarial+Networks

4. Zhu, J., Kaplan, R., Johnson, J., & Fei-Fei, L. (2018). HiDDeN: Hiding data with deep networks. In

Proceedings of the European Conference on Computer Vision (ECCV) (pp. 657–672).

https://scholar.google.com/scholar?q=HiDDeN+Hiding+Data+with+Deep+Networks

Page 702

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

5. Zhang, K. A., Cuesta-Infante, A., Xu, L., & Veeramachaneni, K. (2019). SteganoGAN: High capacity

video steganography with GANs. arXiv preprint arXiv:1901.03892.

https://scholar.google.com/scholar?q=SteganoGAN+High+Capacity+Image+Steganography+with+GAN

6. Boroumand, M., Chen, M., & Fridrich, J. (2019). Deep residual network for steganalysis of digital images.

IEEE Transactions on Information Forensics and Security, 14(5), 1181–1193.

https://scholar.google.com/scholar?q=Deep+Residual+Network+for+Steganalysis+of+Digital+Images

7. Tabares-Soto, R., Orozco-Arias, S., Romero-Cano, V., Bucheli, V. S., & Jiménez-Varón, C. F. (2019).

Deep learning applied to steganalysis of digital images: A systematic review. IEEE Access, 7, 68970–

68990.

https://scholar.google.com/scholar?q=Deep+Learning+Applied+to+Steganalysis+of+Digital+Images+A

+Systematic+Review

8. Duan, J., Jia, D., & Zhao, C. (2019). Reversible video steganography based on U-Net architecture. IEEE

Access, 7, 160728–160742.

https://scholar.google.com/scholar?q=Reversible+Image+Steganography+Based+on+U-

Net+Architecture

9. Chen, Y., Wang, H., & Liu, X. (2020). High-capacity robust video steganography based on adversarial

networks. KSII Transactions on Internet and Information Systems, 14(8), 3382–3398.

https://scholar.google.com/scholar?q=High-

Capacity+Robust+Image+Steganography+Based+on+Adversarial+Networks

10. Chen, Y., Xing, Z., & Liu, X. (2020). Technology of hiding and protecting secret image based on two-

channel deep hiding network. IEEE Access, 8, 30325–30335.

https://scholar.google.com/scholar?q=Technology+of+Hiding+and+Protecting+Secret+Image+Based+o

n+Two-Channel+Deep+Hiding+Network

11. Zhou, X., Zhang, Y., & Wang, H. (2020). Security enhancement of video steganography using generative

adversarial networks. IEEE Signal Processing Letters, 27, 1660–1664.

https://scholar.google.com/scholar?q=Security+Enhancement+of+Image+Steganography+Using+Genera

tive+Adversarial+Networks

12. Manjula, G. R., & Sushma, R. B. (2021). Video steganography: A survey of techniques and methodologies.

In Smart Data Intelligence 2021 (pp. 1–11). Springer, Singapore.

https://scholar.google.com/scholar?q=Video+Steganography+A+Survey+of+Techniques+and+Methodol

ogies

13. Li, X., Wang, J., & Zhang, Y. (2021). Secure video steganography via generative adversarial networks.

EURASIP Journal on Image and Video Processing, 2021(1), 1–15.

https://scholar.google.com/scholar?q=Secure+Image+Steganography+via+Generative+Adversarial+Net

works

14. Wang, Y., & Chen, B. (2022). DWT–DCT based secure video steganography for multimedia

communication. In Proceedings of the IEEE International Conference on Communication and Signal

Processing (pp. 233–237).

https://scholar.google.com/scholar?q=DWT-

DCT+Based+Secure+Video+Steganography+for+Multimedia+Communication

15. Wani, A., Sharma, S., & Gupta, R. (2022). Deep learning-based video steganography: A review. WIREs

Data Mining and Knowledge Discovery, 12(6), e1481.

https://scholar.google.com/scholar?q=Deep+Learning-Based+Image+Steganography+A+Review

16. Rahman, M., Islam, M., & Hasan, M. (2022). Adaptive neural network-based video steganography for

secure communication. Multimedia Tools and Applications, 81(24), 34987–35008.

https://scholar.google.com/scholar?q=Adaptive+Neural+Network-

Based+Image+Steganography+for+Secure+Communication

17. Zhou, Y., Liu, J., & Wang, X. (2022). Encrypted video steganography using deep neural networks. IEEE

Access, 10, 44281–44293.

https://scholar.google.com/scholar?q=Encrypted+Image+Steganography+Using+Deep+Neural+Network

18. Mou, C., Xu, Y., Song, J., Zhao, C., Ghanem, B., & Zhang, J. (2023). Large-capacity and flexible video

steganography via invertible neural network. In Proceedings of the IEEE/CVF Conference on Computer

Page 703

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

Vision and Pattern Recognition Workshops (pp. 1–10).

https://scholar.google.com/scholar?q=Large-

Capacity+and+Flexible+Video+Steganography+via+Invertible+Neural+Network

19. Wang, H., Li, X., & Zhang, Y. (2023). Vision transformer-based video steganography for secure

multimedia communication. Multimedia Tools and Applications, 82(15), 22847–22866.

https://scholar.google.com/scholar?q=Vision+Transformer-

Based+Image+Steganography+for+Secure+Multimedia+Communication

20. Liu, Z., Chen, Y., & Zhao, H. (2023). Video transformer hiding framework for secure video

steganography. Signal Processing: Image Communication, 112, 116934.

https://scholar.google.com/scholar?q=Video+Transformer+Hiding+Framework+for+Secure+Video+Steg

anography

21. Sousa, R. T., Silva, F., & Oliveira, L. (2024). Stego-STFAN: Spatial–temporal feature aggregation

network for secure video steganography. Journal of Information Security and Applications, 78, 103612.

https://scholar.google.com/scholar?q=Stego-STFAN+Spatial-

Temporal+Feature+Aggregation+Network+for+Secure+Video+Steganography

22. Alhamdani, A., Hassan, M., & Ahmed, R. (2025). DeepSteg: Deep learning-based video steganography

with object tracking. Multimedia Systems, 31(2), 1–18.

https://scholar.google.com/scholar?q=DeepSteg+Deep+Learning-

Based+Video+Steganography+with+Object+Tracking

23. Cheddad, A., Condell, J., Curran, K., & McKevitt, P. (2010). Digital video steganography: Survey and

analysis of current methods. Signal Processing, 90(3), 727–752.

https://doi.org/10.1016/j.sigpro.2009.08.010

24. Fridrich, J., & Kodovský, J. (2012). Rich models for steganalysis of digital images. IEEE Transactions on

Information Forensics and Security, 7(3), 868–882.

https://doi.org/10.1109/TIFS.2012.2190402

25. Kheddar, H., Hemis, M., Himeur, Y., Megías, D., & Amira, A. (2024). Deep learning for steganalysis

of diverse data types: A review of methods, taxonomy, challenges and future directions.

Neurocomputing, 587, 127588.

https://doi.org/10.1016/j.neucom.2024.127588

26. Liu, S., Zhang, H., Zhao, Y., & Wang, J. (2023). Video steganography: Recent advances and

challenges. Multimedia Tools and Applications, 82(27), 41943–41985.

https://doi.org/10.1007/s11042-023-15158-0

27. Luo, X., Liu, F., Lian, S., et al. (2021). Deep learning-based steganography and steganalysis: A survey.

ACM Computing Surveys, 54(2), 1–36.

https://doi.org/10.1145/3437479

28. Weng, X., Li, Y., Chi, L., & Mu, Y. (2018). Convolutional video steganography with temporal residual

modeling. In L. Leal-Taixé & S. Roth (Eds.), ECCV 2018 Workshops (LNCS Vol. 11134, pp. 237–

253). Springer, Cham.

https://doi.org/10.1007/978-3-030-11021-5_18

29. Zhang, R., Zhu, H., Liu, F., & Liu, J. (2024). Video steganography based on deep convolutional neural

networks. Multimedia Tools and Applications, 83(4), 10357–10378.

https://doi.org/10.1007/s11042-023-17158-6