Bio-Inspired Quantum-Resistant

Steganographic Communication System

Project Report (PROJ-IT781)
Submitted in partial fulfillment of the requirements for the degree of

Bachelor of Technology (Information Technology)

Team Members:
ANKUR GOSWAMI (13000222013)
KESHAB DEY (13000222019)
ATUL KUMAR SINGH (13000222050)
ASHES DAS (13000222052)

Supervisor:
Dr. SUBHAMITA MUKHERJEE

Department of Information Technology

Techno Main Salt Lake
Kolkata - 700091

December 11, 2025

 CERTIFICATE

This is to certify that the project entitled “ Bio-Inspired Quantum-Resistant Steganographic

Communication System” is a bonafide record of work carried out by ANKUR GOSWAMI,
KESHAB DEY, ATUL KUMAR SINGH, and ASHES DAS under my supervision as part of the
requirements for the degree of Bachelor of Technology in Information Technology at Techno
Main Salt Lake.

Full Signature of the Candidates (with date)

1. Date:

2. Date:

3. Date:

4. Date:

Supervisor: Date:

Head of Department: Date:

Acknowledgement

We would like to express our sincere gratitude to the Department of Information Technology,
Techno Main Salt Lake, and to our project supervisor Dr. Subhamita Mukherjee for constant
guidance and encouragement. We also thank faculty members, peers, friends, and family for
their support throughout the project.
Abstract

Modern secure communication systems face two converging threats: the practical risk of
steganographic detection and the long-term cryptographic risk posed by quantum computing.
Most existing solutions address these problems in isolation—traditional cryptography ignores
detectability, and typical steganographic schemes rely on weak payload embedding that fails
against modern [Link] project proposes a pragmatic, multi-layer communication
framework that combines bio-inspired encoding, hybrid classical–post-quantum key protection,
and GAN-based steganography into a single pipeline designed for confidentiality, stealth, and
future-proof resilience. The pipeline begins with a compact DNA-style 3-bit codon encoding
enhanced with salting and positional shifts to remove predictable structure. A custom symmetric
encryption layer is applied to the codon stream to generate a session BioKey, which is then
wrapped inside a hybrid RSA-OAEP and CRYSTALS-Kyber key encapsulation step to ensure
backward compatibility with existing systems while providing quantum-resistant protection.
The final encrypted payload is embedded into a cover image using an adaptive GAN-based
encoder capable of distributing hidden data with significantly reduced statistical footprints
compared to classical LSB-based methods. This project focuses on developing a functional
prototype of the full end-to-end pipeline, validating correctness through reversible decoding
tests, and evaluating security using PSNR/SSIM metrics and baseline steganalysis tools.
Contents

Acknowledgement ii

Abstract iii

List of Symbols, Abbreviations and Nomenclature viii

Chapter 1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Scope of the Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Organization of the Report . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

Chapter 2 Literature Review 5

Chapter 3 Problem Definition and System Modules 8

3.1 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.1.1 Key Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 Project Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.3 System Modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3.1 Module 1: Bio Inspired Encoder . . . . . . . . . . . . . . . . . . . . . 10
3.3.2 Module 2: Custom Symmetric Encryption Layer . . . . . . . . . . . . 10
3.3.3 Module 3: Hybrid RSA and Kyber Key Wrap . . . . . . . . . . . . . . 10
3.3.4 Module 4: GAN Based Steganographic Embedding . . . . . . . . . . . 10
3.3.5 Module 5: Decoder and Reconstruction Module . . . . . . . . . . . . . 11
3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

Chapter 4 Software Design 12

4.1 Design Philosophy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2.1 Architecture Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.3 Flow of Data Through the System . . . . . . . . . . . . . . . . . . . . . . . . 14
4.3.1 Stage 1: Plaintext Input and Optional Lossless Compression . . . . . . 14
4.3.2 Stage 2: Codon Encoding with Salting and Positional Shift . . . . . . . 14

iv
 4.3.3 Stage 3: DNA-Based Symmetric Encryption . . . . . . . . . . . . . . . 14
4.3.4 Stage 4: RSA Wrap Followed by Kyber PQC Wrap . . . . . . . . . . . 15
4.3.5 Stage 5: Packaging and Optional Compression . . . . . . . . . . . . . 15
4.3.6 Stage 6: GAN-Based Adaptive Embedding . . . . . . . . . . . . . . . 15
4.4 GAN-Based Steganography Module . . . . . . . . . . . . . . . . . . . . . . . 15
4.4.1 GAN Architecture Overview . . . . . . . . . . . . . . . . . . . . . . . 15
4.4.2 Encoder Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.4.3 Decoder Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.4.4 Critic (Discriminator) . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.4.5 Loss Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.4.6 Implementation Details . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.4.7 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.4.8 Security Enhancements . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.4.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.5 Receiver Side Pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.6 Sequence Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.7 Module Interaction Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.8 Design Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

Chapter 5 Software and Hardware Requirements 19

5.1 Hardware Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.1.1 Minimum Hardware Requirements . . . . . . . . . . . . . . . . . . . . 20
5.1.2 Recommended Hardware Requirements . . . . . . . . . . . . . . . . . 20
5.2 Software Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.2.1 Operating System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.2.2 Programming Languages and Runtime . . . . . . . . . . . . . . . . . . 21
5.2.3 Python Libraries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.2.4 Additional Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.3 Dataset Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.4 Cryptographic Requirement Environment . . . . . . . . . . . . . . . . . . . . 21
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

Chapter 6 Implementation 23
6.1 Technology Stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
6.2 Bio-Inspired Codon Encoder . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
6.3 Custom Symmetric DNA Encryption Layer . . . . . . . . . . . . . . . . . . . 25
6.4 Hybrid RSA–Kyber Key Wrap . . . . . . . . . . . . . . . . . . . . . . . . . . 25
6.5 Packaging and Serialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

v
 6.6 GAN-Based Steganographic Embedding . . . . . . . . . . . . . . . . . . . . . 26
6.7 End-to-End Pipeline Integration . . . . . . . . . . . . . . . . . . . . . . . . . 28
6.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

Chapter 7 Experimental Results and Analysis 29

7.1 Cryptographic and Encoding Validation . . . . . . . . . . . . . . . . . . . . . 30
7.1.1 Reversibility Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
7.1.2 Hybrid RSA–Kyber Wrap Performance . . . . . . . . . . . . . . . . . 30
7.2 GAN Steganography Experimental Analysis . . . . . . . . . . . . . . . . . . . 30
7.2.1 Visual Quality: PSNR and MSE . . . . . . . . . . . . . . . . . . . . . 31
7.2.2 Payload Reconstruction Accuracy . . . . . . . . . . . . . . . . . . . . 31
7.2.3 Embedding Heatmap Visualization . . . . . . . . . . . . . . . . . . . 32
7.2.4 Training Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
7.3 End-to-End Pipeline Verification . . . . . . . . . . . . . . . . . . . . . . . . . 32
7.3.1 Steganalysis Resistance . . . . . . . . . . . . . . . . . . . . . . . . . 33
7.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
7.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

Chapter 8 Conclusion and Future Work 34

8.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
8.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
8.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

References 36

List of Tables

7.1 Timing analysis for hybrid key wrapping and unwrapping. . . . . . . . . . . . . 30

7.2 Visual distortion metrics for GAN-based embedding. . . . . . . . . . . . . . . 31
7.3 Steganalysis detection accuracy on GAN stego images. . . . . . . . . . . . . . 33

List of Figures

4.1 System Architecture of this Project . . . . . . . . . . . . . . . . . . . . . . . . 14

vi
4.2 High-Level Sequence of Operations . . . . . . . . . . . . . . . . . . . . . . . 17
7.1 Training convergence for bit recovery accuracy. The red dashed line marks the
beginning of robust fine-tuning. . . . . . . . . . . . . . . . . . . . . . . . . . . 31
7.2 Embedding visualization showing (left) original cover, (middle) stego image,
and (right) residual noise amplified ×50. . . . . . . . . . . . . . . . . . . . . . 32
7.3 Loss minimization curve showing both data and image losses. The dashed
vertical line indicates the start of robust training with quantization noise injection. 32

vii
List of Symbols, Abbreviations and Nomenclature

RSA Rivest–Shamir–Adleman cryptosystem

PQC Post-Quantum Cryptography

GAN Generative Adversarial Network

PSNR Peak Signal-to-Noise Ratio

SSIM Structural Similarity Index Measure

viii
CHAPTER 1

Introduction
Secure communication systems today operate under two simultaneous pressures: rapidly ad-
vancing cryptanalysis capabilities and increasingly sophisticated detection mechanisms that
expose the presence of hidden data. Classical public-key cryptosystems such as RSA, which are
fundamental to secure communication infrastructures, are vulnerable to quantum algorithms
like Shor’s algorithm. In parallel, traditional steganographic methods—particularly LSB-based
techniques—are routinely detected by modern steganalysis tools, making covert communication
significantly harder. As a result, confidentiality alone is no longer sufficient; communication
systems must also minimize detectability and maintain long-term security in a future where
quantum attacks become practical.
This Project is motivated by this gap between encryption strength, forward security, and
stealth. The project explores a practical, modular approach that integrates three distinct layers:
(1) a bio-inspired data encoding stage that transforms plaintext into compact DNA-like codons,
(2) a hybrid cryptographic layer that uses a custom symmetric cipher combined with RSA-
OAEP and CRYSTALS-Kyber to provide both classical and post-quantum key protection, and
(3) a GAN-based steganographic encoder that embeds the encrypted payload into images with
reduced statistical visibility. These individual techniques exist in literature, but their integration
into a single, operational pipeline is largely unexplored. This Project does not attempt to
propose unbreakable cryptography; instead, it focuses on constructing a realistic, experimental
framework that can be evaluated and improved upon.
From an engineering standpoint, the project aims to build a proof-of-concept system im-
plemented in Python, with an emphasis on reversibility, correctness, and baseline security
evaluation rather than claiming industrial-grade cryptographic assurance. The custom sym-
metric cipher is treated explicitly as exploratory, with no claim of formal security proofs. Its
purpose is to examine whether lightweight codon-level transformations can add meaningful
diversity to a multilayer pipeline without introducing excessive complexity. Meanwhile, the hy-
brid RSA–Kyber wrapping ensures that even if the custom cipher is weak, the system maintains
strong cryptographic guarantees under both classical and quantum threat models.
Steganography serves as the final layer of defense. Instead of relying on predictable bit-level
embedding, the system employs a GAN-based approach that learns how to distribute payloads in
image regions that minimize detectable distortions. While training a high-performance GAN is
computationally expensive, the project focuses on integrating available models and measuring
payload impact using PSNR, SSIM, and simple steganalysis benchmarks. The emphasis is on
understanding practical limitations rather than overstating stealth capabilities.

1.1 Motivation

The motivation for this work arises from three converging realities:

1. Quantum threats are no longer theoretical. Once scalable quantum computers emerge,

2
 widely deployed algorithms such as RSA and ECC will be broken, compromising long-
term confidentiality of stored or intercepted data.

2. Steganography is becoming easier to detect. Modern neural-network-based steganalysis

can detect classical embedding schemes with high accuracy, reducing the effectiveness
of traditional covert communication.

3. Research lacks integrated solutions. Existing works focus separately on DNA encoding,
PQC, or GAN steganography, but few attempt to construct a unified, testable pipeline that
combines all three.

1.2 Objectives

The project pursues the following measurable objectives:

• Develop a reversible bio-inspired encoding scheme using 3-bit codons with salting and
positional shifts.

• Implement an experimental symmetric cipher and generate a session BioKey for each
message.

• Secure the BioKey using a hybrid RSA-OAEP and CRYSTALS-Kyber mechanism to

ensure post-quantum resistance.

• Integrate a GAN-based steganographic encoder capable of embedding encrypted payloads

into images with acceptable distortion metrics.

• Validate system correctness through reversible decoding tests and evaluate quality through
PSNR, SSIM, and baseline steganalysis tools.

1.3 Scope of the Project

The scope of this Project is intentionally constrained to ensure feasibility:

• The custom symmetric cipher is exploratory and not presented as a secure alternative to
AES or other standardized ciphers.

• The GAN component relies on available models or lightweight retraining due to hardware
limitations.

• The system is developed as a CLI and prototype API; full-scale UI or mobile applications
are outside the first-phase scope.

• Performance optimization is not a primary goal; correctness, integration, and evaluation

take priority.

3
Despite these constraints, the system is designed to demonstrate the viability of a layered,
quantum-aware, covert communication pipeline.

1.4 Organization of the Report

The remainder of this report is organized as follows:

• Chapter 2 reviews existing research in DNA cryptography, classical cryptography, post-

quantum cryptography, and GAN-based steganography.

• Chapter 3 defines the problem in detail and describes the modules of the proposed
system.

• Chapter 4 presents the software design, pipeline flow diagrams, and architectural com-
ponents.

• Chapter 5 outlines the software and hardware requirements.

• Chapter 6 describes key code structures and implementation approaches.

• Chapter 7 provides experimental results and output samples.

• Chapter 8 concludes the report and discusses potential improvements and future work.

4
 CHAPTER 2

Literature Review
 DNA cryptography models information encoding using nucleotide bases (A, T, C, G), with
early approaches relying on simple mappings and biological metaphors. Although these meth-
ods offer conceptual advantages such as high-density storage and natural obfuscation, practical
implementations suffer from data expansion issues and lack formal cryptographic rigor [1],
[2]. Many schemes rely on simplistic 2-bit mappings that leave statistical fingerprints, while
ASCII-based representations introduce significant data growth.

Classical cryptography such as RSA remains foundational due to mature mathematical proper-
ties and widespread adoption [3]. However, it faces fatal vulnerabilities to quantum algorithms,
particularly Shor’s algorithm, which makes the migration toward quantum-resistant schemes
critically important [4]. Modern best practices employ hybrid encryption that combines efficient
symmetric encryption with asymmetric key-wrapping mechanisms.

Post-quantum cryptography addresses quantum threats through standardized schemes devel-

oped under the NIST PQC initiative [5]. Lattice-based algorithms such as CRYSTALS-Kyber
provide strong security guarantees and competitive performance but introduce larger key sizes
and greater computational overhead [6]. Practical implementations offered by the Open Quan-
tum Safe Project enable researchers to integrate PQC algorithms experimentally into software
systems [7]. Recent developments in quantum steganography demonstrate the feasibility of
quantum-resistant techniques in IoT environments, including reversible decoding and cus-
tomized encryption for secure smart city communication [8]. Despite these advances, inte-
gration complexity and resource constraints remain significant challenges.

Traditional image steganography techniques are computationally inexpensive but remain vul-
nerable to statistical detection methods [9]. GAN-based steganography significantly improves
imperceptibility and embedding quality through adversarial learning and neural feature op-
timization [10], [11]. Systems such as SteganoGAN demonstrate practical, high-capacity
embedding using generative models [12]. However, these techniques require substantial train-
ing resources and carefully curated datasets. Advanced watermarking schemes further extend
robustness through multi-level embedding strategies [13].

Modern deep learning architectures particularly StyleGAN variants have dramatically im-
proved image generation fidelity, providing robust foundations for generative steganography
and adaptive cover construction [10]. These capabilities enable more sophisticated embedding
strategies that resist contemporary detection frameworks. Meanwhile, ongoing research in
quantum steganography continues to investigate hybrid models that combine quantum-resistant
cryptography with traditional embedding mechanisms for distributed IoT systems requiring si-
multaneous secrecy and stealth [8].

6
Implementation frameworks such as PyCryptodome supply essential cryptographic primitives
including hashing, encryption, secure randomness, and authenticated modes that streamline
practical system development [14]. These libraries allow researchers to focus on architectural
and algorithmic innovation instead of low-level cryptographic engineering.

The existing literature reveals a significant gap in fully integrated systems that combine multiple
cryptographic layers with formal evaluation. DNA-based steganography frequently omits PQC
integration, while GAN-based steganography lacks strong internal cryptographic design. Quan-
tum steganography bridges some gaps but remains centered on IoT-specific constraints rather
than general-purpose secure communication. This project aims to address these limitations
by developing a unified framework that incorporates bio-inspired encoding, quantum-resistant
cryptography, and advanced steganographic techniques for broad, real-world applicability.

7
 CHAPTER 3

Problem Definition and System Modules

Modern secure communication depends on two properties. The first is cryptographic confi-
dentiality, which prevents an attacker from understanding the message. The second is stealth,
which prevents an attacker from even detecting that a message exists. Failure in either property
can compromise the entire communication process. Classical cryptosystems such as RSA are
vulnerable to future quantum attacks, while traditional steganographic techniques are increas-
ingly detectable by machine learning based steganalysis. Existing research focuses on these
areas separately, but practical systems rarely combine strong encryption, post-quantum safety,
and adaptive steganography into a unified design.

3.1 Problem Definition

The core problem addressed in this project is the creation of an end-to-end communication
pipeline that can securely transform plaintext into an encrypted, quantum-resistant, and covert
payload that remains statistically inconspicuous when embedded inside an image. The objective
is not to claim mathematically proven cryptographic strength for experimental components,
but to design a system in which the failure of any single layer does not compromise overall
confidentiality.

3.1.1 Key Challenges

• Classical public key cryptography becomes insecure in the presence of scalable quantum
computers, which can run Shor’s algorithm and break RSA.

• Traditional LSB based image steganography produces predictable statistical artifacts that
are easily detectable by modern steganalysis tools.

• DNA based data encoding schemes seen in literature usually lack strong security proofs
and often create large data expansion.

• Multi layer systems require careful integration. The output of one stage must remain
compatible and efficiently embeddable in subsequent stages.

• The system must remain reversible under all conditions. Any error in encoding, encryp-
tion, or embedding should not break decryption.

3.2 Project Goals

This project aims to engineer a functional, layered communication system that satisfies the
following goals:

• Provide confidentiality that remains safe against both classical and anticipated quantum
attackers.

9
 • Reduce detectability by embedding encrypted payloads in images using adaptive machine
learning based steganography.

• Use bio inspired codon encoding to add an additional reversible obfuscation layer before
cryptographic processing.

• Combine RSA and CRYSTALS Kyber to produce a hybrid key wrapping mechanism.

• Implement a pipeline where each stage can be tested independently and as part of the
complete system.

3.3 System Modules

The proposed system is divided into five major modules. Each module is designed to perform
one specific task so that testing and debugging remain manageable.

3.3.1 Module 1: Bio Inspired Encoder

This module converts plaintext into binary and then into fixed length 3 bit codons. Random
salting and positional shifts are applied to remove deterministic output patterns. The output is
a codon sequence that is compact and reversible.

3.3.2 Module 2: Custom Symmetric Encryption Layer

A lightweight symmetric cipher is applied to the codon stream. The purpose of this cipher is
exploratory. It is not intended to replace AES or other standardized algorithms. It generates a
session BioKey that is later protected by stronger cryptographic layers.

3.3.3 Module 3: Hybrid RSA and Kyber Key Wrap

The BioKey is encrypted twice. First with RSA OAEP for classical compatibility. Then with
CRYSTALS Kyber for post quantum safety. This hybrid approach ensures that the system
remains secure even if one cryptographic assumption fails.

3.3.4 Module 4: GAN Based Steganographic Embedding

A generative adversarial network model is used to embed the final encrypted package into a cover
image. Unlike classical LSB methods, the GAN selects embedding locations that minimize
detectable distortions. The goal is to reduce the risk of detection by standard steganalysis tools.

10
3.3.5 Module 5: Decoder and Reconstruction Module

On the receiver side, the pipeline is reversed. The system extracts the hidden payload from the
stego image, unwraps the key using Kyber and RSA, decrypts the codon stream, and restores
the original plaintext. Complete reversibility is a primary requirement for validating system
correctness.

3.4 Summary

The problem addressed by this Project arises from the combined threat of quantum attacks
and advanced steganalysis. Existing techniques do not fully solve this problem when used
individually. This chapter defined the overall system goals and described each functional
module. The next chapter presents the design diagrams and the pipeline architecture used to
integrate these modules.

11
CHAPTER 4

Software Design
This chapter describes the software design of this Project .The system uses a sequential, multi-
layer pipeline where data flows through encoding, encryption, key wrapping, packaging, and
GAN-based embedding. Unlike traditional web-oriented systems, this project emphasizes
algorithmic and data-flow design rather than DFDs or ER diagrams. Each block in the pipeline
corresponds directly to a processing stage in the architecture.

4.1 Design Philosophy

The design follows three core principles:

• Each module performs one clearly defined function with minimal shared state.

• The output of every stage must be a valid, verifiable, and reversible input for the next.

• Compromise or failure in any single stage should not compromise total confidentiality.

4.2 System Architecture

The system architecture (Figure 4.1) follows a sender-to-receiver pipeline. The sender trans-
forms plaintext into a hidden stego image using sequential processing layers. The receiver
performs the reverse operations.

• Plaintext / File Input

• Optional Lossless Compression

• Codon Encoding + Salting/Shift

• DNA-Based Symmetric Encryption (BioKey generation)

• RSA Wrap → PQC (Kyber) Wrap

• Packaging and Optional Compression

• GAN-Based Adaptive Embedding

• Stego Image Output

13
4.2.1 Architecture Diagram

Figure 4.1: System Architecture of this Project

4.3 Flow of Data Through the System

4.3.1 Stage 1: Plaintext Input and Optional Lossless Compression

The system accepts text, binary files, or arbitrary payloads. Optional lossless compression
reduces payload size and improves embedding efficiency.

4.3.2 Stage 2: Codon Encoding with Salting and Positional Shift

The payload is converted into 3-bit codons. Salting and positional shifting introduce controlled
randomness to prevent structural patterns.

4.3.3 Stage 3: DNA-Based Symmetric Encryption

The codon stream is encrypted using a lightweight symmetric cipher, producing the session key
known as the BioKey. Later RSA and Kyber layers secure this key.

14
4.3.4 Stage 4: RSA Wrap Followed by Kyber PQC Wrap

1. The BioKey is first wrapped using RSA.

2. The RSA output is then encapsulated using CRYSTALS-Kyber.

This hybrid scheme provides both classical and quantum-resistant security.

4.3.5 Stage 5: Packaging and Optional Compression

The encrypted codon payload, RSA-wrapped key, Kyber-wrapped key, and metadata are com-
bined into a binary package.

4.3.6 Stage 6: GAN-Based Adaptive Embedding

Since the GAN embedding module is the most complex component, its design is described in
more detail below.

4.4 GAN-Based Steganography Module

4.4.1 GAN Architecture Overview

The GAN steganography system consists of three networks:

• Encoder: Embeds the payload into the cover image.

• Decoder: Recovers the hidden payload.

• Critic: Enforces realism in the stego output.

4.4.2 Encoder Structure

The Encoder receives a cover image 𝐶 and payload 𝑆 as a 4-channel tensor:

𝐶 ′ = 𝐸 (𝐶, 𝑆)

4.4.3 Decoder Structure

𝑆′ = 𝐷𝑒𝑐(𝐶 ′)

Training minimizes bit error rate.

4.4.4 Critic (Discriminator)

The Critic outputs the probability that an image is natural. It provides the adversarial signal
that improves stealth.

15
4.4.5 Loss Functions

The total loss balances payload accuracy, low distortion, and adversarial realism:

𝐿 total = 𝜆data 𝐿 data + 𝜆 img 𝐿 img + 𝜆 adv 𝐿 adv

1. Data Loss (bit reconstruction, BCE).

𝑁
1 ∑︁ 
𝑠𝑖 log(𝑠𝑖′) + (1 − 𝑠𝑖 ) log(1 − 𝑠𝑖′)

𝐿 data =−
𝑁 𝑖=1

2. Image Loss (MSE for fidelity).

𝐻 𝑊
1 ∑︁ ∑︁
𝐿 img = ∥𝐶 (𝑥, 𝑦) − 𝐶 ′ (𝑥, 𝑦)∥ 2
𝐻𝑊 𝑥=1 𝑦=1

3. Adversarial Loss (statistical realism).

   
𝐿 adv = E𝐶∼𝑃cover ℓD (𝐷 (𝐶), 1) + E𝐶,𝑆 ℓD (𝐷 (𝐸 (𝐶, 𝑆)), 0)

Weighting constants.
𝜆 data = 50, 𝜆 img = 1, 𝜆 adv = 0.05

Robustness is improved using quantization noise:

 
𝑆′robust 1 1


= 𝐷𝑒𝑐 𝐸 (𝐶, 𝑆) + 𝑁 , 𝑁 ∼𝑈 − 255 , 255

4.4.6 Implementation Details

Training uses COCO Test2017. The optimizer is Adam with 𝛼 = 0.0001, batch size 32 and the
losses defined above.

4.4.7 Experimental Results

• Bit Accuracy: 99.9985% (raw)

• Bit Accuracy with Reed–Solomon: 100%

• PSNR: 29.06 dB

• MSE: 0.00497

• Payload Density: less than 0.5 bpp

Residual analysis shows high-frequency embedding improves stealth.

16
4.4.8 Security Enhancements

• Reed–Solomon correction improves robustness.

• Adaptive payload density prevents over-embedding.

4.4.9 Conclusion

The GAN module provides a robust and stealthy embedding mechanism aligned with reversibil-
ity and confidentiality requirements.

4.5 Receiver Side Pipeline

1. GAN decoder extracts hidden data.

2. Unpacking retrieves encrypted codons and wrapped keys.

3. Kyber unwrap retrieves RSA output; RSA unwrap retrieves the BioKey.

4. The codon stream is decrypted using the BioKey.

5. Salting and shifting are reversed to restore plaintext.

4.6 Sequence Diagram

Figure 4.2: High-Level Sequence of Operations

4.7 Module Interaction Summary

• Encoding produces salted codons.

17
 • Encryption yields the encrypted stream and BioKey.

• RSA + Kyber wrapping protects the key.

• Packaging prepares data for embedding.

• GAN embedding hides the package inside the image.

4.8 Design Limitations

• GAN performance depends heavily on data quality.

• The symmetric cipher is experimental.

• Embedding capacity is limited by perceptual distortion thresholds.

• GAN + PQC add computational overhead.

4.9 Conclusion

The software design integrates bio-inspired encoding, hybrid cryptography, and adaptive GAN-
based embedding into a reversible, modular, and secure pipeline.

18
 CHAPTER 5

Software and Hardware Requirements

This chapter outlines the hardware and software environment required to develop, train, and
execute the components of this Project. Since the system integrates cryptographic processing,
GAN-based steganography, and Python-based tooling, the requirements span both computa-
tional and development domains.

5.1 Hardware Requirements

5.1.1 Minimum Hardware Requirements

• Processor: Dual-core CPU (Intel i3 / Ryzen 3 or higher)

• RAM: 8 GB

• Storage: 10 GB free disk space

• GPU: Not required for inference-only execution

These minimum specifications are sufficient for running encryption, encoding, and basic
steganographic decoding. GAN inference at reduced speed is also feasible.

5.1.2 Recommended Hardware Requirements

• Processor: Quad-core CPU (Intel i5 / Ryzen 5 or higher)

• RAM: 16 GB or higher

• GPU: NVIDIA GPU with CUDA support (6 GB VRAM recommended)

• Storage: SSD with 50+ GB free space for datasets and models

These specifications enable full training of the GAN steganography model and fast execution
of cryptographic routines such as Kyber encapsulation.

5.2 Software Requirements

5.2.1 Operating System

• Windows 10 / 11

• Ubuntu 20.04+ (recommended for training)

• macOS (for non-GPU workflows)

20
5.2.2 Programming Languages and Runtime

• Python 3.10+

• [Link] (optional for auxiliary scripts)

• CUDA toolkit (for GPU training)

5.2.3 Python Libraries

• PyTorch — GAN training and inference

• TorchVision — image preprocessing and augmentation

• PyCryptodome — symmetric and RSA cryptography

• liboqs-python — CRYSTALS-Kyber PQC primitives

• NumPy, SciPy — numerical operations

• Pillow — image processing

5.2.4 Additional Tools

• Jupyter Notebook (model experimentation)

• Google Colab / Kaggle Notebooks (cloud GPU training)

• Git (version control)

• Overleaf (documentation preparation)

5.3 Dataset Requirements

The GAN model used in this Project is trained on the COCO Test2017 dataset:

• 40,000 natural images

• Diverse scenes for better embedding generalization

• Suitable for adaptive texture-based payload hiding

5.4 Cryptographic Requirement Environment

RSA and Kyber require strong cryptographic randomness:

• High-entropy system RNG (/dev/urandom on Linux, OS RNG on Windows)

• Secure key storage for RSA and Kyber keypairs

21
5.5 Conclusion

The above hardware and software stack provides the necessary foundation for implementing,
training, and evaluating the complete this project pipeline. While the system can run in a CPU-
only environment, full GAN training performance benefits significantly from a CUDA-capable
GPU.

22
CHAPTER 6

Implementation
This chapter describes the practical implementation of this Project .The system follows a modular
architecture, where each processing stage—encoding, encryption, key wrapping, packaging, and
GAN-based embedding which is implemented as an independent component. This ensures clean
separation of responsibilities, easier debugging, and end-to-end reversibility.

6.1 Technology Stack

The implementation is developed in Python 3.10. The major libraries used are:

• PyCryptodome — RSA encryption, hashing, and utility primitives.

• liboqs-python — CRYSTALS-Kyber post-quantum key encapsulation.

• NumPy, Pillow — data manipulation and image processing.

• PyTorch — implementation of GAN encoder, decoder, and critic.

• TorchVision — dataset loading and augmentation.

The codebase is organized into modular packages:

Project/
|-- encoder/ # Codon encoder + inverse decoder
|-- crypto/ # DNA cipher, RSA wrap, Kyber wrap
|-- packaging/ # Serialization, metadata, compression
|-- gan/ # Encoder, decoder, critic, robust training
|-- utils/ # Helpers, logging, configuration
|-- [Link] # End-to-end pipeline runner

6.2 Bio-Inspired Codon Encoder

The implementation begins by converting plaintext into a compact stream of 3-bit codons. A
random salt and positional shift are applied to eliminate deterministic output patterns.

Procedure

1. Convert plaintext bytes to a raw bitstream.

2. Pad the bitstream to a multiple of 3.

3. Group bits into 3-bit segments.

4. Apply salt-based XOR and positional shifts.

5. Output salted, obfuscated codon sequence.

24
Pseudocode

function encode_to_codons(data):
bits = to_binary(data)
padded = pad_to_multiple(bits, 3)
codons = split_into_triplets(padded)
salted = xor_with_salt(codons)
shifted = apply_positional_shift(salted)
return shifted

The decoder performs the inverse procedure to restore plaintext.

6.3 Custom Symmetric DNA Encryption Layer

A lightweight symmetric cipher encrypts the codon stream. Although experimental, this cipher
enables analysis of bio-inspired encryption behavior. The cipher generates a session key known
as the BioKey.

Operation

The cipher uses XOR substitution and cyclic rotations. It is not intended as a replacement for
standardized ciphers (AES), but is fully protected later by RSA and Kyber.

function dna_encrypt(codon_stream):
key = generate_random_key()
enc = xor_rotate(codon_stream, key)
return enc, key

Decryption applies inverse XOR and reverse rotations.

6.4 Hybrid RSA–Kyber Key Wrap

To achieve quantum-resistant confidentiality, the BioKey is wrapped using a layered hybrid

scheme:

1. RSA-OAEP encrypts the BioKey (classical security).

2. CRYSTALS-Kyber encapsulates the RSA ciphertext (post-quantum security).

25
Implementation Steps

function hybrid_wrap(bioKey):
rsa_cipher = RSA_OAEP_encrypt(bioKey)
pqc_cipher, kem_key = Kyber_encapsulate(rsa_cipher)
return pqc_cipher, kem_key
Unwrapping performs the reverse:
rsa_plain = Kyber_decapsulate(pqc_cipher)
bioKey = RSA_OAEP_decrypt(rsa_plain)
This approach ensures that even if RSA becomes vulnerable, the PQC layer still protects the
key.

6.5 Packaging and Serialization

The encrypted codon stream, RSA wrap, Kyber wrap, salts, and metadata are bundled into a
compact binary structure.

Package Format

{
"encCodon": <encrypted codon stream>,
"rsaWrap" : <RSA ciphertext>,
"kyberWrap": <PQC ciphertext>,
"meta": {
"salt": ...,
"shiftParam": ...,
"length": ...
}
}
The packager optionally applies compression to improve embedding efficiency.

6.6 GAN-Based Steganographic Embedding

The most computationally intensive component is the GAN steganography model. It imple-
ments:
• Encoder: embeds the package into a cover image.

• Decoder: extracts the package from the stego image.

• Critic: adversarial discriminator that enforces visual realism.

26
Training Setup

• Dataset: COCO Test2017 (40,000 images)

• Optimizer: Adam, 𝛼 = 0.0001

• Batch size: 32

• Loss weights: 𝜆 𝑑𝑎𝑡𝑎 = 50, 𝜆𝑖𝑚𝑔 = 1, 𝜆 𝑎𝑑𝑣 = 0.05

Robust Training for PNG Safety

A robustness phase injects uniform noise:

 
1 1
𝑁 ∼𝑈 − ,
255 255
into the stego image before decoding. This ensures 100% payload recovery even after integer
quantization (PNG compression).

Embedding Procedure

Function SecureEmbed(coverImage, message):

H, W ← Dimensions(coverImage)
coverTensor ← Normalize(coverImage)
dataBytes ← UTF8_Encode(message)
bits ← BytesToBits(dataBytes)
IF Length(bits) > H * W THEN
Return Error("Message too large")
END IF
Pad bits with zeros to size H * W
payload ← Reshape(bits, shape=(1, H, W))
inputTensor ← Concatenate(coverTensor, payload)
stegoTensor ← Encoder_Network(inputTensor)
stegoImage ← Quantize(stegoTensor)
Return stegoImage
End Function

Extraction is the reverse:

Function SecureExtract(stegoImage):
tensor ← Normalize(stegoImage)
raw ← Decoder_Network(tensor)
bits ← Threshold(raw, 0.5)

27
 bytes ← BitsToBytes(bits)
message ← UTF8_Decode(bytes)
Return message
End Function

(This entire GAN module is evaluated experimentally in Chapter 7.)

6.7 End-to-End Pipeline Integration

The complete sender pipeline is implemented as:

function sender(message, coverImage):

codons = encode_to_codons(message)
encCodon, bioKey = dna_encrypt(codons)
rsaWrap, pqcWrap = hybrid_wrap(bioKey)
packageBlob = pack(encCodon, rsaWrap, pqcWrap)
stego = embed(packageBlob, coverImage)
return stego

The receiver pipeline executes the reverse sequence:

function receiver(stegoImage):
package = extract(stegoImage)
encCodon, wraps = unpack(package)
bioKey = hybrid_unwrap(wraps)
codons = dna_decrypt(encCodon, bioKey)
plaintext = decode_codons(codons)
return plaintext

6.8 Summary

The implementation achieves a fully functional prototype of a quantum-resistant, bio-inspired

steganographic communication system. Each module is independently testable, and the in-
tegration pipeline strictly follows the architecture described earlier. The GAN model adds
a strong layer of adaptive, low-visibility data embedding, completing the end-to-end secure
communication framework.

28
 CHAPTER 7

Experimental Results and Analysis

This chapter presents the experimental evaluation of this Project. The system combines bio-
inspired encoding, multi-layer hybrid cryptography, and GAN-based adaptive steganography.
Each component was tested individually and as part of the complete pipeline to validate cor-
rectness, robustness, stealth, and reversibility.
The results are organized into three major domains:

• Cryptographic correctness and reversibility

• Steganographic performance (PSNR, MSE, bit recovery)

• GAN model behavior (training dynamics, embedding patterns, robustness)

7.1 Cryptographic and Encoding Validation

7.1.1 Reversibility Test

A corpus of 200 payloads ranging from 32 bytes to 32 KB was passed through the full sender–
receiver pipeline. All outputs matched their corresponding inputs exactly.

• Reconstruction Accuracy: 100%

• Bit Errors: 0 across all tests

This confirms the correctness of codon encoding, symmetric encryption, positional shifting,
and hybrid key wrapping.

7.1.2 Hybrid RSA–Kyber Wrap Performance

The hybrid cryptographic module was evaluated for latency and reliability.

Operation Average Time (ms) Success Rate

RSA-OAEP Encryption 1.82 100%

Kyber-1024 Encapsulation 3.41 100%
RSA-OAEP Decryption 1.74 100%
Kyber-1024 Decapsulation 3.29 100%

Table 7.1: Timing analysis for hybrid key wrapping and unwrapping.

The system meets both classical and post-quantum cryptographic requirements.

7.2 GAN Steganography Experimental Analysis

The GAN steganography module was evaluated in terms of visual quality, payload recovery,
training convergence, and embedding behavior.

30
7.2.1 Visual Quality: PSNR and MSE

A set of 100 images from COCO Test2017 were used to measure distortion after embedding.

Metric Value Interpretation

PSNR 29.06 dB High imperceptibility

MSE 0.00497 Low pixel distortion
SSIM 0.92 Strong structural similarity

Table 7.2: Visual distortion metrics for GAN-based embedding.

7.2.2 Payload Reconstruction Accuracy

The decoder’s bit recovery accuracy was recorded across training epochs. A major jump in
accuracy occurs after fine-tuning with quantization robustness.

Figure 7.1: Training convergence for bit recovery accuracy. The red dashed line marks the
beginning of robust fine-tuning.

Final results:

• Raw Bit Accuracy: 99.9985%

• Bit Accuracy with Reed–Solomon: 100%

Reed–Solomon correction ensures perfect payload reconstruction even after PNG quantization.

31
7.2.3 Embedding Heatmap Visualization

The embedding pattern was analyzed by visualizing the residual noise between the cover and
stego images. The GAN correctly concentrates modifications in complex textures such as hair,
foliage, and shadows, while preserving flat regions such as skin or the sky.

Figure 7.2: Embedding visualization showing (left) original cover, (middle) stego image, and
(right) residual noise amplified ×50.

The residual confirms adaptive, texture-aware embedding—reducing detectability.

7.2.4 Training Dynamics

Loss values (data loss, image loss) were tracked during training using a logarithmic scale. A
sharp improvement is observed after robustness fine-tuning.

Figure 7.3: Loss minimization curve showing both data and image losses. The dashed vertical
line indicates the start of robust training with quantization noise injection.

7.3 End-to-End Pipeline Verification

End-to-end testing validates the complete pipeline.

32
 • End-to-End Success Rate: 100%

• Sender Runtime: 185 ms (avg.)

• Receiver Runtime: 162 ms (avg.)

7.3.1 Steganalysis Resistance

Classical detectors were applied to stego images.

Detector Detection Accuracy Interpretation

Chi-square Test 51.2% Random chance (good)

RS Analysis 49.8% Random chance (good)
Sample Pairs Analysis 50.7% No detectable bias

Table 7.3: Steganalysis detection accuracy on GAN stego images.

This indicates strong resistance against standard statistical attacks.

7.4 Discussion

The combined experimental evaluation demonstrates that:

• The cryptographic pipeline is fully reversible with zero data loss.

• GAN-based embedding introduces minimal visual distortion.

• Payload recovery is perfect after Reed–Solomon correction.

• Embedding heatmaps confirm adaptive, texture-aware hiding.

• The model is robust against PNG quantization.

• Classical steganalysis tools fail to detect hidden content.

These results validate the effectiveness of the integrated system for confidential, covert commu-
nication.

7.5 Conclusion

This Project successfully demonstrates that bio-inspired encoding, hybrid cryptography, and
GAN-based steganography can be combined into a robust and future-resistant communication
pipeline. The system achieves strong confidentiality, high reversibility, and low detectability,
positioning it as a promising approach for next-generation secure communication systems.

33
 CHAPTER 8

Conclusion and Future Work

8.1 Conclusion

This Project presents a complete, end-to-end framework for secure and covert digital commu-
nication. The system integrates bio-inspired 3-bit codon encoding, a lightweight symmetric
DNA-style cipher, hybrid RSA–Kyber key protection, and a GAN-based adaptive stegano-
graphic encoder. Each stage is fully reversible and modular, enabling independent testing and
systematic debugging.
The experimental evaluation demonstrates that the GAN module achieves near-perfect data
recovery (99.9985% raw accuracy and 100% with Reed–Solomon correction), producing stego
images whose distortions remain below perceptual thresholds (PSNR 29 dB, MSE 0.00497).
The hybrid key wrapping successfully protects the BioKey under both classical and post-
quantum threat models, satisfying the confidentiality requirements of the system.

8.2 Limitations

While the system meets its design goals, certain practical limitations remain:

• The custom symmetric cipher is experimental and not mathematically validated.

• GAN-based embedding requires GPU acceleration for practical training and inference.

• Embedding capacity is constrained to maintain steganographic invisibility.

• PQC operations such as Kyber encapsulation add significant computational overhead.

8.3 Future Work

The system can be extended in several directions:

• Replace the experimental symmetric cipher with NIST-standardized lightweight ciphers.

• Explore diffusion-model-based steganography for higher imperceptibility.

• Integrate error-correcting codes more deeply for noisy-channel environments.

• Support video-based or audio-based steganography for multi-modal secure communica-

tion.

• Develop a GUI-based desktop or web application for real-world usability.

Overall,this Project demonstrates a viable direction for next-generation secure communication

systems that must remain stealthy, confidential, and resilient against both classical and quantum
adversaries.

35
References

[1] M. A. Tabatabaei and A. A. H. Fahmi, “Dna-based cryptography: A survey,” Journal of

Information Security and Applications, vol. 45, pp. 64–83, 2018.
[2] M. Abe and J. Emerick, “Dna steganography and cryptography: A review and comparative
survey,” arXiv preprint arXiv:1905.08264, 2019, arXiv:1905.08264.
[3] J. Katz and Y. Lindell, Introduction to Modern Cryptography, 2nd. CRC Press, 2014.
[4] P. W. Shor, “Algorithms for quantum computation: Discrete logarithms and factoring,” in
Proceedings of the 35th Annual Symposium on Foundations of Computer Science, 1994,
pp. 124–134.
[5] National Institute of Standards and Technology, Post-quantum cryptography, https:
//[Link]/projects/post-quantum-cryptography, 2024.
[6] D. J. Bernstein, J. Buchmann, and E. Dahmen, Eds., Post-Quantum Cryptography.
Springer Verlag Berlin Heidelberg, 2009.
[7] Open Quantum Safe Project, Liboqs, [Link] 2024.
[8] A. Kumar, S. Sharma, and P. Gupta, “Advancing quantum steganography: A secure iot
communication with reversible decoding and customized encryption technique for smart
cities,” Cluster Computing, vol. 27, no. 8, pp. 10 245–10 268, 2024.
[9] J. Fridrich, M. Goljan, and R. Du, “Detecting lsb steganography in color and gray-scale
images,” IEEE Multimedia, vol. 8, no. 4, pp. 22–28, 2001.
[10] T. Karras et al., “Analyzing and improving the image quality of stylegan,” in Proceed-
ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020,
pp. 8110–8119.
[11] M. Volkhonskiy, S. V. Denisov, and M. Maj, Steganographic generative adversarial
networks, arXiv preprint, 2019.
[12] Steganogan - dai lab, [Link] 2020.
[13] M. Barni, F. Bartolini, R. Caldelli, and I. Cox, “Multilevel watermarking and steganog-
raphy,” IEEE Signal Processing Magazine, vol. 21, no. 6, pp. 31–45, 2004.
[14] Pycryptodome - the python cryptography toolkit, [Link]

36