Exploring Generative Adversarial Networks for Secure Data

Encryption and Future Directions in Communication Systems

Ranjith Bhat

1,2

and Raghu Nanjundegowda

Dept of Robotics and AI Engineering, NMAM Institute of Technology, NITTE (Deemed to be University) Nitte, India

Dept of Electronics and Communication JAIN (Deemed to be University) Bengaluru, India

Dept of Electrical and Electronics Engineering JAIN (Deemed to be University) Bengaluru, India

Keywords: Artificial Intelligence, Generative Adversarial Network (GANs), Image Encryption, Multilevel GAN.

Abstract: The rapid advancements in communication systems and the proliferation of digital technologies have

underscored the critical need for robust and adaptive encryption methods to safeguard data integrity,

confidentiality, and authenticity. Traditional cryptographic techniques, while effective, face challenges in the

wake of evolving cyber threats and emerging technologies such as quantum computing. This paper explores

the transformative potential of Generative Adversarial Networks (GANs) in secure data encryption and

communication systems. By leveraging the dynamic architecture of GANs, which consists of a generator and

a discriminator operating in an adversarial framework, novel encryption methodologies are developed. These

methodologies address limitations in traditional encryption by introducing non-linear, adaptive encryption

schemes resistant to reverse engineering and capable of generating dynamic encryption keys. The paper

further investigates the integration of GANs into modern communication paradigms, including quantum

communication, blockchain networks, and IoT systems. Additionally, it highlights the challenges in adopting

GAN-based encryption, including training instability, scalability, and adversarial vulnerabilities, while

proposing solutions to overcome these issues. Through experimental validation, the study demonstrates the

superior security and efficiency of GAN-based encryption systems, offering a scalable and intelligent

approach to securing data in an increasingly complex digital landscape.

1 INTRODUCTION

The rapid advancement of communication systems

and the proliferation of digital technologies have

fundamentally reshaped the way data is exchanged,

stored, and processed (Goodfellow, 2014). From

personal communications to global financial systems,

the reliance on secure data transmission has become

an indispensable requirement in ensuring the

integrity, confidentiality, and authenticity of

information (Cao, 2020). However, the ever-

increasing sophistication of cyberattacks and data

breaches has exposed the vulnerabilities in existing

encryption methodologies, demanding more robust

and adaptive solutions for securing communication

systems. Traditional cryptographic techniques such

as symmetric and asymmetric encryption methods

have been the cornerstone of secure communication

https://orcid.org/0009-0000-6149-1243

https://orcid.org/0000-0002-2091-8922

for decades. While these methods are effective

against many contemporary threats, the rise of

quantum computing and other disruptive

technologies presents significant challenges to their

long-term viability (Wang, 2018). As attackers

develop more advanced techniques, it becomes

imperative to explore innovative and intelligent

approaches to encryption that can not only resist these

threats but also adapt to evolving attack vectors in real

time. This has led researchers to investigate the

potential of emerging technologies, such as artificial

intelligence (AI) and machine learning (ML), to

revolutionize secure communication (Singh, 2023).

Generative Adversarial Networks (GANs)

represent a significant advancement in artificial

intelligence and machine learning, functioning as an

effective mechanism for the generation of realistic

synthetic data, including images, videos, and text.

292

Bhat, R. and Nanjundegowda, R.

Exploring Generative Adversarial Networks for Secure Data Encryption and Future Directions in Communication Systems.

DOI: 10.5220/0013591300004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 292-300

ISBN: 978-989-758-763-4

Since their introduction by Ian Goodfellow and

colleagues in 2014, Generative Adversarial Networks

(GANs) have been extensively utilized in various

applications, such as image synthesis, data

augmentation, and anomaly detection, among others

(Li, 2020). Generative Adversarial Networks (GANs)

fundamentally comprise two neural networks: a

generator and a discriminator. These networks engage

in a competitive process characterized as a zero-sum

game, as illustrated in Fig. 1. The adversarial dynamic

enables GANs to learn intricate data distributions and

produce outputs that cannot be differentiated from

real data (Zahmoul, 2016).

Figure 1: A typical GAN architecture.

The training set consists of the set of real images

illustrated in the figure, while random noise is

provided to the Generator to initiate the training

process. Both the neural networks are trained further

and updated as per the loss function available through

the back propagation. In the context of secure data

encryption, GANs offer an intriguing paradigm shift

(Li and Li, 2019). Unlike traditional encryption

methods that rely on deterministic algorithms, GANs

can generate highly complex and adaptive encryption

schemes that are inherently resistant to reverse

engineering. By leveraging their ability to learn and

adapt, GANs can create non-linear, dynamic

encryption keys that are extremely difficult for

adversaries to decipher, even with access to advanced

computational resources. Additionally, GANs can be

utilized to detect and counteract security threats in

real time, further enhancing the resilience of

communication systems (Bhat and Nanjundegowda,

2025).

This paper explores the potential of GANs in

transforming secure data encryption and

communication systems. We propose a novel

framework for leveraging GANs to develop adaptive

encryption algorithms that can address the limitations

of traditional methods while offering enhanced

protection against emerging threats. Furthermore, we

investigate the integration of GANs into futuristic

communication paradigms, such as quantum

communication systems, blockchain-based networks,

and the Internet of Things (IoT), where the need for

innovative security solutions is paramount (Zhang,

2020).

In addition to discussing the strengths and

potential of GAN-based encryption, we also examine

the challenges and limitations associated with their

adoption. Issues such as computational complexity,

scalability, and the risk of adversarial attacks on

GANs themselves are critical factors that must be

addressed to realize their full potential (Zhang, 2018).

Furthermore, ethical considerations and regulatory

frameworks for deploying AI-driven encryption

techniques will be explored, ensuring that these

technologies are implemented responsibly and

securely (Zhao, 2022).

The structure of this paper is as follows: Section 2

provides a comprehensive overview of GANs, their

architecture, and key principles. Section 3 delves into

the application of GANs for secure data encryption,

outlining proposed methodologies and use cases.

Section 4 discusses the challenges, limitations, and

potential risks associated with GAN-based

encryption systems. Section 5 highlights future

research directions and opportunities for advancing

GANs in the context of secure communication.

Finally, Section 6 concludes the paper by

summarizing key findings and emphasizing the

transformative potential of GANs in redefining

secure communication systems.

Through this research, we aim to bridge the gap

between cutting-edge AI technologies and the

pressing need for advanced encryption mechanisms,

providing a foundation for future innovations in

secure communication (Li, 2019). By harnessing the

power of GANs, we envision a new era of adaptive,

resilient, and intelligent communication systems

capable of withstanding the challenges of an

increasingly complex digital landscape (Bhat, 2025).

2 RELATED WORKS

A neural network was designed for impulsive

coordination within the reaction-diffusion

mechanism, effectively modelling the dynamic

behaviours of these systems (Chen, 2016). This

method was later utilized for image encryption

purposes. Chaotic systems demonstrate notable

cryptographic potential, particularly in the context of

image cryptosystems, providing robust security

features against various traditional attacks, such as

plaintext attacks. The neural network described was

later employed in image cryptosystems. A scheme

Exploring Generative Adversarial Networks for Secure Data Encryption and Future Directions in Communication Systems

293

that combines chaotic systems with neural networks,

resulting in a solution that exhibits improved security

and decreased complexity compared to previous

methods (Dridi, 2016). An image encryption scheme

utilizing a stacked autoencoder network to generate

chaotic sequences. The scheme exhibited significant

efficiency, attributable to the parallel computing

capabilities of the stacked autoencoder and its

resilience against conventional attacks (Hu, 2017). In

a specific study, a new image steganography

technique that avoided the embedding of messages

within carrier images. The deep model demonstrated

significant improvements in image security metrics,

exhibited an effective extraction phase, and showed

strong resilience against steganalysis algorithms (Hu,

2018).

An image encryption approach (Li, 2018) using a

CNN trained on the CASIA iris dataset (Debiasi,

2015) to generate encryption keys. Iris characteristics

were retrieved and encoded using RS error-correcting

codes. The encoded vector was used to XOR-encrypt

plain images. An encryption keys with a Montgomery

County chest X-ray dataset-trained GAN (Ding,

2020). This updated system has a larger key space,

better pseudo-randomness, resilience to typical image

processing assaults, and higher modification

sensitivity (Jaeger, 2014). A scheme utilizing a deep

neural network that removed the requirement for pre-

shared keys between systems (Jin, 2020). The system

dynamically generated and utilized encryption keys,

resulting in enhanced overall security. A DNN-based

image encryption scheme that employs the SIPI

image dataset. This scheme integrates chaotic maps

for the encryption process, ensuring the preservation

of image quality (Manivath, 2020).

An encryption method that employs multiple

chaotic sequences generated from sensitive keys,

which were derived by training a convolutional neural

network (CNN) on the ImageNet database (Erkan,

2022). The initial conditions for encryption in the

hyperchaotic logistic map were determined using

parameters produced by the network. A two-layer

deep neural network aimed at classifying silica

aerogel (SA) in the context of physical unclonable

functions (Fratalocchi, 2020). The chaotic behavior

of SA was employed to produce cryptographic keys,

yielding random key sequences for various input

conditions. We strongly encourage authors to use this

document for the preparation of the camera-ready.

Please follow the instructions closely in order to make

the volume look as uniform as possible (Moore and

Lopes, 1999).

An image encryption scheme that employs a

Cycle-GAN architecture. The network was trained

using a dataset consisting of both plain and cipher

satellite images. This approach utilized double

random phase encoding to achieve image encryption

(Li, 2021). An alternative scheme utilizing Cycle-

GAN, which was trained on a chest X-ray dataset

(Ding, 2021). This scheme not only executed

encryption-decryption tasks but also detected specific

objects within the cipher images. The flaws in prior

techniques to establish a foundation for an improved

avalanche impact (Bao, 2021). A sophisticated

framework was introduced that integrates a diffusion

mechanism. The neural network, trained on satellite

image datasets from Google Maps, demonstrated

enhanced efficiency; nonetheless, it displayed

inadequate performance in the decryption process

(Baluja, 2017).

Cycle-GAN networks are extensively utilized in

encryption and decryption operations inside deep

learning-based image encryption frameworks,

including picture steganography, showcasing their

versatility in modern cryptographic applications. An

experimental results reveal that CryptoGAN achieves

high levels of randomness and unpredictability,

essential for secure encryption, and provides strong

resistance to cryptanalysis (Bhat, 2024). This study

highlights the potential of CryptoGAN to

revolutionize image security by combining traditional

cryptography with advanced machine learning

techniques. At addressing limitations in traditional

encryption methods like AES and chaotic encryption,

CryptoGAN combines U-Net as the generator and

PatchGAN as the discriminator to encrypt and

decrypt images while maintaining high visual fidelity

and robust security (Bhat, 2024). Trained on a dataset

of 2000 butterfly images, CryptoGAN ensures

structural similarity, high entropy, and low pixel

correlation, effectively resisting cryptanalysis and

statistical attacks. The model achieves superior

performance compared to existing methods, with high

SSIM and PSNR values.

3 CHALLENGES

Advanced GAN-based models have received

significant focus in the field of cybersecurity.

Nonetheless, the implementation of these methods in

encryption and decryption presents certain

challenges. This section examines the primary

challenges faced in utilizing GANs for cybersecurity,

with a focus on protecting digital assets and the

effective implementation of these models. The use of

GANs in encryption and decryption faces several

technical challenges, including training instability

INCOFT 2025 - International Conference on Futuristic Technology

294

and mode collapse, which may adversely affect the

performance and reliability of these models. The

integration of GANs into existing security

frameworks presents a significant challenge,

necessitating precise alignment to guarantee seamless

functionality and scalability.

3.1 Cryptographic Challenges

Despite their potential, GANs face notable challenges

when applied to image data encryption. One

significant issue is the inconsistency and limited

diversity in the quality of generated image data. This

limitation can undermine the effectiveness of using

GAN-generated images in testing encryption

algorithms, where reliability and diversity are

essential. Additionally, GAN training can be

unstable, often leading to difficulties in achieving

convergence. This instability not only complicates the

optimization process but also impacts the evaluation

of the model's performance in encryption-related

tasks.

To address these challenges, a new symmetric

encryption framework called Adversarial Neural

Cryptography (ANC) has been introduced,

specifically designed for image data. ANC integrates

GANs into its structure to enhance encryption

capabilities and provide robust security against

chosen-ciphertext attacks (CCA). The ANC system

models secure communication between two entities,

Alice and Bob, who exchange encrypted image data

using a shared key K. Meanwhile, Eve, a passive

attacker, attempts to decode the plaintext image P by

analysing the ciphertext C.

In developing the ANC system, particular focus is

given to resisting CCA attacks. The system employs

a multi-layer encryption strategy coupled with a

sophisticated key exchange mechanism to minimize

the statistical correlation between plaintext images

and their corresponding ciphertext. This approach

significantly increases the difficulty for attackers

attempting to breach the system. Additionally, the

GAN's generator is utilized to introduce higher levels

of randomness and unpredictability to the ciphertext

images, further bolstering the system's resilience to

CCA attacks.

Experimental findings confirm the effectiveness

of ANC in mitigating CCA threats. By leveraging the

GAN's ability to enhance the randomness in

ciphertext, the feasibility of attackers conducting

statistical analysis is greatly reduced. Fig. 2 illustrates

the symmetric encryption and decryption model used

in ANC for image data using 2 parties Alice and Bob

using the same symmetric Key which is both used for

Encryption and Decryption. The experiments also

evaluate ANC's performance in simulated attack

scenarios, highlighting its robustness in protecting

encrypted image communication and ensuring secure

exchanges. Overall, while challenges such as training

instability and variability in data quality exist, the

integration of GANs into cryptographic systems like

ANC demonstrates their transformative potential. By

addressing these limitations, ANC effectively

harnesses GANs to improve both the efficiency and

security of encryption methodologies for image data,

paving the way for more advanced and secure

cryptographic systems.

Figure 2: Symmetric Encryption scheme used by two

parties

3.2 Cybersecurity Challenges

This section highlights key challenges in

cybersecurity, particularly in addressing adversarial

attacks and implementing advanced techniques like

GANs and federated learning.

3.2.1 Adversarial Attacks and Adversarial

Example Generation

Challenges posed by adversarial evasion attacks,

where altered input samples deceive classifiers,

compromising botnet detection accuracy (Randhawa,

2021). While efforts to improve recall rates and

address dataset imbalance using GANs for synthetic

oversampling show promise, challenges remain in

generating diverse, high-quality datasets and keeping

up with evolving attack methods. Further, modern

botnets require updated traffic features for effective

differentiation, emphasizing the need for continuous

research.

The emergence of Adversarial Examples (AEs) in

cybersecurity. AEs are malicious perturbations that

mislead classifiers, posing threats to machine learning

(ML)-based systems (Zhang, 2020). While most

research on AEs focuses on computer vision, their

impact on cybersecurity systems remains

underexplored, underscoring the need for robust ML

models that can withstand adversarial attacks.

Challenges in countering adversarial attacks,

where malicious samples deceive both humans and

ML systems (Schneider, 2023). These attacks exploit

Exploring Generative Adversarial Networks for Secure Data Encryption and Future Directions in Communication Systems

295

vulnerabilities in malware classifiers and pose

significant risks to cybersecurity (Lucas, 2023).

Defence strategies, like adversarial training and

frameworks such as Défense-GAN, aim to enhance

robustness against such attacks, but their

effectiveness varies across datasets and attack types

(Laykaviriniyakul, 2023).

3.3 Network Security Challenges

The growing challenges in network security, driven

by rapid technological advancements and an

expanding number of internet users (Yang, 2022).

The increase in network traffic, fuelled by the rise of

5G networks, and the emergence of threats like trojan

horses, viruses, and phishing sites have made

detecting and mitigating network threats more

complex. This necessitates improved methods for

proactive defence and network threat detection.

In a related study, (Das, 2022) emphasized the

challenges posed by the dynamic nature of computer

and mobile networks. Increasing nodes and traffic

complicate anomaly detection and adaptation to

modern attacks. Privacy concerns in intrusion

detection systems and challenges like coordinating

updates in large-scale networks and preventing model

tampering were addressed using federated learning.

This approach enables secure sharing of encrypted

models while preserving data privacy.

The vulnerabilities arising from diversified access

points in 5G and distributed networks, which have

expanded the attack surface (Park, 2022). The

increasing frequency and sophistication of

cyberattacks make detection and prevention more

difficult, emphasizing the need for enhanced intrusion

detection systems to safeguard networks. Limitations

in current botnet detection methods, noting their

inability to fully capture the evolving and

sophisticated behaviours of botnets (Yin, 2018).

These adaptive threats, which leverage advanced

technologies to evade detection, present a significant

challenge, underscoring the need for more

comprehensive network flow analysis.

4 SUGGESTED

METHODOLOGIES FOR

ENCRYPTION WITH GANS

Designing a robust image encryption scheme using

Generative Adversarial Networks (GANs) involves

leveraging the unique architecture of GANs, which

consists of a generator and a discriminator. These two

neural networks operate in a competitive framework

where the generator produces encrypted versions of

images, and the discriminator evaluates the

authenticity or quality of these outputs. This

adversarial process allows the generator to learn

intricate transformations that obscure the content of

the original image while maintaining a structured

framework for decryption.

It is assumed that both the generator (G) and

discriminator (D) models possess sufficient capacity

to handle the required tasks. When the generator's

data distribution 𝑝



(𝑥) aligns perfectly with the real

data distribution 𝑝



(𝑥), the GAN model achieves

a state of equilibrium. At this point, the discriminator

D cannot distinguish between real and generated data,

resulting in a classification accuracy of 50%. Here,

𝑝



(𝑥) represents the distribution of data generated by

the generator. Formally, for a specific generator G,

the optimal discriminator D* can be determined.

A commonly used approach in GANs is the

hierarchical structure, which allows encrypted images

to be generated step-by-step, gradually improving

their resolution at each stage. This hierarchical

architecture is particularly beneficial for applications

that require high-quality outputs, such as image

encryption. For instance, MultiLevelGAN utilizes

this method to generate progressively detailed

outputs, as depicted in Fig. 3. Compared to traditional

encryption techniques, GANs demonstrate a

significant advantage in terms of generation speed.

By replacing the traditional sampling process with a

generator, GANs eliminate the need for a lower

bound to approximate likelihood, streamlining the

generation process.

A critical component of the encryption process is

the generation of secure and pseudo-random keys.

GANs can be trained to generate such keys by

learning from chaotic systems like logistic maps,

which provide high randomness and unpredictability.

The generator produces encryption keys that are

inherently complex and difficult to decipher, ensuring

the robustness of the encryption process.

INCOFT 2025 - International Conference on Futuristic Technology

296

Figure 3: Architecture of MultiLevelGAN.

These keys form the foundation for encryption

operations, including substitution, permutation, and

diffusion, which collectively transform the original

image into an unintelligible form. Substitution

modifies the pixel values based on the generated key,

permutation rearranges the pixel positions to disrupt

spatial coherence, and diffusion ensures that small

changes in the original image result in significant

differences in the encrypted output.

The encryption process begins with training the

GAN using a dataset of images, where the generator

learns to encrypt the images, and the discriminator

assesses the quality of encryption. The goal of

training is for the generator to produce encrypted

images that are indistinguishable from a target

distribution, effectively confusing the discriminator.

This iterative adversarial training ensures that the

generator develops the capability to perform highly

secure and adaptive encryption. The discriminator, in

turn, becomes a robust evaluator of the encryption

quality, pushing the generator to continually improve.

Once the encryption process is established, the

decryption mechanism reverses the transformations

applied during encryption. Using the same key

generated by the GAN, the encrypted image

undergoes inverse diffusion, permutation, and

substitution to reconstruct the original image. The

decryption process is designed to be lossless,

ensuring that the original image is retrieved without

any degradation in quality. This reversibility is a

critical aspect of the encryption scheme, as it ensures

usability without compromising security.

The security of the GAN-based encryption

scheme is rigorously analysed to confirm its

robustness. Statistical analysis is performed on the

encrypted images to verify the uniformity of pixel

value distributions, indicating effective encryption.

Key sensitivity analysis ensures that even slight

variations in the key render the decryption process

ineffective, highlighting the system’s dependency on

the exact key for secure operations. Additionally, the

scheme is subjected to various attacks, including

brute force, differential, and statistical attacks, to

evaluate its resilience. Studies have demonstrated that

GAN-based encryption methods are highly resistant

to such attacks, offering a robust framework for

secure image transmission and storage.

Implementing a GAN-based encryption scheme

requires careful consideration of computational

resources and dataset quality. Training GANs is

computationally intensive and demands substantial

processing power. The quality and diversity of the

training dataset significantly influence the GAN’s

ability to generate effective encryption keys.

Furthermore, hyperparameters such as learning rates

and network architectures must be carefully tuned to

achieve optimal performance. Despite these

challenges, GAN-based encryption provides a

flexible and adaptive framework for securing images

in a variety of applications.

In Figure 4, a Secure Transformation Network

(STN) processes plaintext and keys by first

converting them into angles using f(b) as input to the

neural network. The weight matrix multiplication in

the adversarial encrypting network is then computed

to generate the initial ciphertext. The final ciphertext

is obtained by applying the inverse transformation

𝑓



(𝑎)

Notably, all data handled by the STN are

floating-point numbers, with ciphertext values

constrained to the range [0, 1].

Mathematically, the fully connected layer of the

cipher set performs operations as described in

Equation (1):

(

ℎ



ℎ



ℎ



….ℎ



)



(

𝑎



….𝑎



𝑎



….ℎ



)

𝑊 (1)

Here, W represents the unified weight matrix of all

hidden and convolutional layers in the adversarial

encrypting network.

( 𝑎



….𝑎



𝑎



….𝑎



)

corresponds to the

angles of the plaintext and key, while

( ℎ



ℎ



ℎ



….ℎ



)

represents the network's

output variables.

In the rest of the experiment, the cipher set is

expressed mathematically as shown in Equation (2):

This section must be in two columns.

𝐶𝜉(𝑊,𝑃𝐼,𝐾𝐼) (2)

where P, K, and C denote the plaintext, key, and

ciphertext as n-bit vectors, respectively.

Exploring Generative Adversarial Networks for Secure Data Encryption and Future Directions in Communication Systems

297

Figure 4: Neural Network of MultiLevelGAN

This idea presents an analysis of the encryption

structure, algorithm functionality, and security

performance of the Adversarial Neural Cryptography

(ANC) system, specifically when applied to image

data. While ANC has shown potential, previous

research highlights vulnerabilities when ANC is

combined with multi-layer neural networks for

computer communication systems. Specifically, it

has been observed that such systems can be cracked

by adversarial neural networks through training.

To address these challenges, this study proposes

an enhanced adversarial encryption algorithm called

CCA-ANC, tailored for image data. The core idea

behind CCA-ANC is to simulate a stronger attacker

with greater cracking capabilities, thereby forcing the

sender and legitimate receiver to adopt a more robust

encryption system. This approach results in a highly

secure and resilient encryption method.

4.1 Concept of CCA-ANC for Image

Data

The Chosen-Ciphertext Attack (CCA) technique in

CCA-ANC allows an attacker to select a sequence of

ciphertexts and analyse the corresponding plaintext or

key information. This method is particularly effective

for evaluating the security of the ANC algorithm. By

applying CCA, potential weaknesses in the

encryption mechanism can be identified and

addressed, leading to algorithm improvements. For

image data, this technique ensures the encryption

system can withstand sophisticated cryptographic

attacks and enhances the system's overall security and

reliability in real-world applications.

4.2 Continuous XOR for Image

Encryption

One of the novel contributions of this experiment is

the extension of the XOR operation to a continuous

space, optimized for image encryption. Traditional

XOR, commonly used in cryptography, is adapted

using a unit circle representation. The experiment

maps binary values (0 and 1) to corresponding angles

(0 and π), enabling a continuous transformation. The

resulting XOR operation becomes the sum of two

angles, making it more suitable for continuous data,

such as image pixels.

The mapping of bit positions to angles is defined by

the following equations:

1. Mapping bit position to angle:

𝑓

(

𝑏

)

arccos

(

12𝑏

)

(3)

Here in (3), f(b) represents the conversion of bit

position b to an angle.

2. Inverse mapping of angle to continuous bits:

𝑓



(

𝑎

)





(



)



(4)

In (4) inverse function transforms the angle back

to its original bit representation.

This continuous XOR operation enables the

encryption of image data in a floating-point space,

making the process more flexible and secure for high-

resolution and complex image datasets.

4.3 Secure Transformation Network

(STN)

To verify the security of the encryption process, a

Secure Transformation Network (STN) is introduced.

The STN, as shown in Figure 4, is designed to

evaluate the robustness of the encryption mechanism

by learning and detecting potential vulnerabilities.

The structure of STN is as follows:

Input Conversion: The plaintext and keys are

transformed into angles using the f(b) mapping,

converting bits into angles for input into the neural

network.

Adversarial Encryption: A weight matrix

multiplication is performed within the adversarial

encryption network to generate the initial ciphertext.

The STN processes all data as floating-point

numbers, with the resulting ciphertext values

constrained to the range [0, 1]. This ensures precision

and adaptability when encrypting and decrypting

image data.

INCOFT 2025 - International Conference on Futuristic Technology

298

5 CONCLUSIONS AND FUTURE

SCOPES

The potential of GAN-based encryption extends

beyond traditional use cases, with opportunities for

integration into real-time systems and cross-modal

encryption tasks. Future research can focus on

developing specialized GAN architectures tailored

for encryption, optimizing real-time performance,

and expanding the scope of encryption to other data

modalities such as video and audio. By addressing

these directions, GANs can revolutionize secure

communication systems, ensuring the confidentiality

and integrity of data in an increasingly interconnected

digital world. The adaptability and learning

capabilities of GANs make them a promising avenue

for advancing encryption methodologies and

overcoming the challenges posed by emerging cyber

threats.

REFERENCES

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,

Warde-Farley, D., Ozair, S., ... & Bengio, Y. (2014).

Generative adversarial nets. Advances in neural

information processing systems, 27.

Cao, W., Mao, Y., & Zhou, Y. (2020). Designing a 2D

infinite collapse map for image encryption. Signal

Processing, 171, 107457.

Wang, X., Zhu, X., & Zhang, Y. (2018). An image

encryption algorithm based on Josephus traversing and

mixed chaotic map. IEEE Access, 6, 23733-23746.

Singh, M., Baranwal, N., Singh, K. N., & Singh, A. K.

(2023). Using GAN-based encryption to secure digital

images with reconstruction through customized super

resolution network. IEEE Transactions on Consumer

Electronics, 70(1), 3977-3984.

Li, Q., Wang, X., Wang, X., Ma, B., Wang, C., Xian, Y., &

Shi, Y. (2020). A novel grayscale image steganography

scheme based on chaos encryption and generative

adversarial networks. IEEE Access, 8, 168166-168176.

Zahmoul, R., & Zaied, M. (2016, October). Toward new

family beta maps for chaotic image encryption. In 2016

IEEE International Conference on Systems, Man, and

Cybernetics (SMC) (pp. 004052-004057). IEEE.

Zhang, Y., Jia, R., Pei, H., Wang, W., Li, B., & Song, D.

(2020). The secret revealer: Generative model-

inversion attacks against deep neural networks. In

Proceedings of the IEEE/CVF conference on computer

vision and pattern recognition (pp. 253-261).

Zhang, R., Isola, P., Efros, A. A., Shechtman, E., & Wang,

O. (2018). The unreasonable effectiveness of deep

features as a perceptual metric. In Proceedings of the

IEEE conference on computer vision and pattern

recognition (pp. 586-595).

Zhao, D., Zhu, X., Liu, B., Ren, J., Zhu, X., Mao, Y., ... &

Ullah, R. (2022). High-security and low-complexity

OCDM transmission scheme based on GAN enhanced

chaotic encryption. Optics Express, 30(19), 34898-

34907.

Li, X., & Li, X. (2019, July). A novel block image

encryption algorithm based on DNA dynamic encoding

and chaotic system. In 2019 IEEE 4th international

conference on signal and image processing (ICSIP) (pp.

901-906). IEEE.

Bhat, R., & Nanjundegowda, R. (2025). A Review on

Comparative Analysis of Generative Adversarial

Networks’ Architectures and Applications. Journal of

Robotics and Control (JRC), 6(1), 53-64.

Chen, W. H., Luo, S., & Zheng, W. X. (2016). Impulsive

synchronization of reaction–diffusion neural networks

with mixed delays and its application to image

encryption. IEEE transactions on neural networks and

learning systems, 27(12), 2696-2710.

Dridi, M., Hajjaji, M. A., Bouallegue, B., & Mtibaa, A.

(2016). Cryptography of medical images based on a

combination between chaotic and neural network. IET

Image Processing, 10(11), 830-839.

Hu, F., Wang, J., Xu, X., Pu, C., & Peng, T. (2017). Batch

image encryption using generated deep features based

on stacked autoencoder network. Mathematical

Problems in Engineering, 2017(1), 3675459.

Hu, D., Wang, L., Jiang, W., Zheng, S., & Li, B. (2018). A

novel image steganography method via deep

convolutional generative adversarial networks. IEEE

access, 6, 38303-38314.

Li, X., Jiang, Y., Chen, M., & Li, F. (2018). Research on

iris image encryption based on deep

learning. EURASIP Journal on Image and Video

Processing, 2018, 1-10.

Debiasi, L., & Uhl, A. (2015, March). Techniques for a

forensic analysis of the casia-iris v4 database. In 3rd

International Workshop on Biometrics and Forensics

(IWBF 2015) (pp. 1-6). IEEE.

Ding, Y., Wu, G., Chen, D., Zhang, N., Gong, L., Cao, M.,

& Qin, Z. (2020). DeepEDN: A deep-learning-based

image encryption and decryption network for internet

of medical things. IEEE Internet of Things

Journal

, 8(3), 1504-1518.

Jaeger, S., Candemir, S., Antani, S., Wáng, Y. X. J., Lu, P.

X., & Thoma, G. (2014). Two public chest X-ray

datasets for computer-aided screening of pulmonary

diseases. Quantitative imaging in medicine and surgery,

4(6), 475.

Jin, J., & Kim, K. (2020). 3D CUBE algorithm for the key

generation method: Applying deep neural network

learning-based. IEEE Access, 8, 33689-33702.

Maniyath, S. R., & Thanikaiselvan, V. (2020). An efficient

image encryption using deep neural network and

chaotic map. Microprocessors and Microsystems, 77,

103134.

Erkan, U., Toktas, A., Enginoğlu, S., Akbacak, E., &

Thanh, D. N. (2022). An image encryption scheme

based on chaotic logarithmic map and key generation

Exploring Generative Adversarial Networks for Secure Data Encryption and Future Directions in Communication Systems

299

using deep CNN. Multimedia Tools and

Applications, 81(5), 7365-7391.

Fratalocchi, A., Fleming, A., Conti, C., & Di Falco, A.

(2020). NIST-certified secure key generation via deep

learning of physical unclonable functions in silica

aerogels. Nanophotonics, 10(1), 457-464.

Li, J., Zhou, J., & Di, X. (2021). A learning optical image

encryption scheme based on CycleGAN. J. Jilin

Univ.(Eng. Technol. Ed.), 51(3), 1060-1066.

Ding, Y., Tan, F., Qin, Z., Cao, M., Choo, K. K. R., & Qin,

Z. (2021). DeepKeyGen: a deep learning-based stream

cipher generator for medical image encryption and

decryption. IEEE Transactions on Neural Networks

and Learning Systems, 33(9), 4915-4929.

Bao, Z., & Xue, R. (2021). Research on the avalanche effect

of image encryption based on the Cycle-GAN. Applied

Optics, 60(18), 5320-5334.

Bao, Z., Xue, R., & Jin, Y. (2021). Image scrambling

adversarial autoencoder based on the asymmetric

encryption. Multimedia Tools and

Applications, 80(18), 28265-28301.

Baluja, S. (2017). Hiding images in plain sight: Deep

steganography. Advances in neural information

processing systems, 30.

Bhat, R., & Nanjundegowda, R. (2024). CryptoGAN: a new

frontier in generative adversarial network-driven image

encryption. Int J Artif Intell, 13(4), 4813-4821.

Bhat, R., & Nanjundegowda, R. (2024). Comparative

Analysis of CryptoGAN: Evaluating Quality Metrics

and Security in GAN-based Image Encryption. Journal

of Robotics and Control (JRC), 5(5), 1557-1569.

Randhawa, R. H., Aslam, N., Alauthman, M., Rafiq, H., &

Comeau, F. (2021). Security hardening of botnet

detectors using generative adversarial networks. IEEE

Access, 9, 78276-78292.

Zhang, S., Xie, X., & Xu, Y. (2020). A brute-force black-

box method to attack machine learning-based systems

in cybersecurity. IEEE Access, 8, 128250-128263.

Schneider, J., & Apruzzese, G. (2023). Dual adversarial

attacks: Fooling humans and classifiers. Journal of

Information Security and Applications, 75, 103502.

Lucas, K., Pai, S., Lin, W., Bauer, L., Reiter, M. K., &

Sharif, M. (2023). Adversarial training for {Raw-

Binary} malware classifiers. In 32nd USENIX Security

Symposium (USENIX Security 23) (pp. 1163-1180).

Laykaviriyakul, P., & Phaisangittisagul, E. (2023).

Collaborative Defense-GAN for protecting adversarial

attacks on classification system. Expert Systems with

Applications, 214, 118957.

Yang, Y., Yao, C., Yang, J., & Yin, K. (2022). A network

security situation element extraction method based on

conditional generative adversarial network and

transformer. IEEE Access, 10, 107416-107430.

Das, S. (2022). FGAN: Federated generative adversarial

networks for anomaly detection in network

traffic. arXiv preprint arXiv:2203.11106.

Park, C., Lee, J., Kim, Y., Park, J. G., Kim, H., & Hong, D.

(2022). An enhanced AI-based network intrusion

detection system using generative adversarial

networks. IEEE Internet of Things Journal, 10(3),

2330-2345.

Yin, C., Zhu, Y., Liu, S., Fei, J., & Zhang, H. (2018, May).

An enhancing framework for botnet detection using

generative adversarial networks. In 2018 International

Conference on Artificial Intelligence and Big Data

(ICAIBD) (pp. 228-234). IEEE.

INCOFT 2025 - International Conference on Futuristic Technology

300