A Secure and Efficient Voice Authentication Framework Based on Frequency Shift Keying Modulation and Modified Optimized RSA Encryption

I. Introduction

II. Literature Review

III. Methodology

Table 1. Summary of system components and their benefits.

IV. System Architecture and Workflow

• Tools: Custom Python MORSA module, GMPY2 for big int

➢

Encryption Algorithm

• Read plaintext message from file or input stream.

• Divide the message into manageable chunks.
• Apply padding to match block size requirements.

• For each chunk:

• Aggregate encrypted chunks into Encrypted Payload.

➢

Decryption Algorithm

• Receive the encrypted payload over the network.

• For each encrypted chunk:

• Unpad and reconstruct message from decrypted chunks.
• Combine segments to form the complete message.

• Save the decrypted message to a file or output interface.

Figure 3. flow diagram for Voice Authentication Framework Combining AI Feature Extraction and MORSA Cryptography.

Figure 3. flow diagram for Voice Authentication Framework Combining AI Feature Extraction and MORSA Cryptography.

➢

Authentication Logic via Cosine Similarity

➢

Blockchain Logging Upon authentication, the following fields are recorded in a blockchain ledger:

●

Timestamp

●

User ID

●

SHA-256 hash of the voiceprint

●

Authentication result (Success/Failure)

●

Cosine similarity score

IV. RESULT AND DISCUSSION

➢

Experimental Evaluation Experiments show that:

●

AI embeddings maintain high cosine similarity (>0.95) with minimal channel noise.

●

MORSA encryption causes negligible distortion when decrypted accurately.

●

Cosine similarity drops with increased Gaussian noise but remains above threshold in low-SNR conditions.

➢

Setup:

●

Consider 100D AI embedding vector

●

Quantized → Encrypted (MORSA) → Gaussian Noise → Decrypted

●

Cosine similarity measured vs original

A. Cosine similarity analysis

Cosine similarity is a metric that calculates the cosine of the angle between two non-zero vectors in an inner product space. It measures directional similarity, not magnitude.

$$\mathit{cos}\theta ={\displaystyle \frac{{\underset{v}{\to }}_{original . }{\underset{v}{\to }}_{reconstructed}}{‖{\underset{v}{\to }}_{original }‖.‖{\underset{v}{\to }}_{reconstructed}‖}}$$

Table 2. Interpretation of Result.

Table 2. Interpretation of Result.

Cosine Similarity	Meaning	Authentication Result
≈ 1.0	Vectors are very similar	Accept (Success)
≈ 0.0	Vectors are orthogonal	Rejection (Failure)
< Threshold	Too different to match	Reject

A threshold (e.g., 0.85) is empirically set to decide success vs failure.

In Figure 4, the line graph compares feature values across feature indices for Original Voiceprint (yellow line with dots) and Reconstructed Voiceprint (thin red line). Here X-axis (Feature Index): Represents the feature indices (0 to 12) extracted from the voiceprint, such as MFCCs or other acoustic features used for speaker identification. And Y-axis (Feature Value): Represents the normalized feature values (0 to 1).

Original Voiceprint (Yellow): The yellow line shows variation across feature indices, with peaks around indices 4, 5, and 11, indicating distinct feature values for the original signal. Reconstructed Voiceprint (Red): The red line is flat at zero, indicating no feature values are present or reconstructed.

Figure 5. Voice feature comparion after FSK transmission and decryption.

Figure 5. Voice feature comparion after FSK transmission and decryption.

The original voiceprint contains significant feature information, as reflected by the fluctuating values across the indices. The reconstructed voiceprint does not contain feature information (all zero), indicating:

• The reconstruction method removed or did not recover voiceprint-relevant features.
• The system successfully prevents voiceprint leakage during reconstruction, enhancing privacy.

This figure demonstrates that the proposed encryption and reconstruction system can preserve signal content for general use while removing sensitive voiceprint features, ensuring privacy-preserving speech transmission.

The experiment shows at fig 5 that as the transmission noise level (σ) increases, the cosine similarity between the original and reconstructed voiceprint vector drops steadily:

• At low noise levels (e.g., σ < 0.02), similarity remains high (≈0.98–1.0), ensuring reliable authentication.
• At moderate noise (σ ≈ 0.05), similarity starts nearing the threshold (e.g., 0.85), risking false rejections.
• At high noise (σ > 0.07), the similarity may fall below the threshold, leading to authentication failure.

MORSA-based systems remain under low-to-moderate noise but require effective channel equalization or error correction for noisy environments.

C. Voice Authentication Using Probabilistic Linear Discriminant Analysis[20]

In voice authentication, PLDA (Probabilistic Linear Discriminant Analysis) is a core scoring technique used to verify if two voice samples belong to the same speaker. It’s often used with i-vector or x-vector speaker embeddings in state-of-the-art systems. The feature is shown in Figure 7.

➢

Pipeline Steps

i.

Voice Feature Extraction :Audio is pre-processed and converted into fixed-length embeddings like I-vectors or x-vectors. Here input is Raw audio and output embedding vector (e.g., 512D).

ii.

PLDA Training : Trains on embeddings from many speakers. Learn to model speaker identity (between-class) and channel/session variation (within-class).

iii.

Scoring : PLDA scores the similarity between a test voice and a reference voice using a log-likelihood ratio (LLR). If the score is above a threshold value means same speaker (accept); otherwise reject.

➢

PLDA Scoring Formula

Given two embedding vectors x₁ and x₂:

$$score\left(x_1,x_2\right)={\displaystyle \frac{P\left({x_{1,}x}_2\right|samespeaker)}{P\left(x_{1,}\right|different\left)P\right(x_{2,}\left|different\right)}}$$

(5.8)

PLDA uses a probabilistic model trained on speaker data to compute this likelihood.

Table 5. Benefit use of PLDA in Voice Authentication.

Table 5. Benefit use of PLDA in Voice Authentication.

Feature	Benefit
Handles channel variability	Reduces false rejections due to different microphones or environments
Probabilistic modelling	Computing a statistically sound match score
Performs well with short utterances	Especially useful in real-world voice authentication
Standard in SOTA systems	Used in NIST SRE and major biometric applications

➢

Tools and Frameworks for Implementation

●

Kaldi Toolkit – Most widely used for PLDA-based speaker verification.

●

Speech Brain / PyTorch-Kaldi – Modern PyTorch-based toolkits.

●

Sci Kit-Learn – For simplified PLDA-like behavior (with limitations).

False Match Rate (FMR) and False Non-Match Rate (FNMR) shown in Table 6. for different voice verification comparators: out of 10,000 attempts by impostors, the system would falsely accept one as a match.

A False Non-Match Rate (FNMR) of 20% means that in a biometric system (like facial recognition), 20% of the time, the system incorrectly rejects a genuine user as a non-match. Essentially, it's the rate at which the system fails to recognize an authorized individual.

The hybrid model effectively integrates deep learning and post-classical cryptography to produce a secure, accurate voice authentication pipeline.

Table 7. FNMR (%) at Different FMR (%) Levels.

Table 7. FNMR (%) at Different FMR (%) Levels.

False Match Rate (%)	False Non-Match Rate (FNMR)%
	Cosine (Plain)	PLDA (Plain)	2Cov (Plain)
0.01	20.0	15.0	12.0
0.1	13.0	9.0	6.0
1.0	8.0	4.5	3.0
5.0	4.0	2.0	1.2
20.0	2.5	1.2	0.8
40.0	2.0	0.9	0.5

D. Comparative Analysis: AI-Based vs. MORSA-Based Voice Authentication Systems

Voice authentication systems have emerged as a critical component in secure identity verification, leveraging unique voiceprint characteristics to authorize access. Two divergent approaches have evolved in this domain: (1) Artificial Intelligence (AI)-based models that utilize machine learning algorithms for speaker recognition, and (2) Cryptographic systems like the Modified RSA (MORSA) cryptosystem, which encrypts biometric data before verification. This section presents a detailed comparison of these paradigms with respect to security, privacy, performance, and deployment feasibility shown in Table 8, Table 9, Table 10, Table 11, Table 12.

Both AI-based and MORSA-based voice authentication systems offer distinct advantages. AI-based systems excel in adaptability, ease of deployment, and accuracy under controlled conditions. However, they are limited by privacy concerns, potential for adversarial attacks, and reliance on probabilistic models.

Conversely, the MORSA-based approach prioritizes security, privacy, and traceability through cryptographic rigor and blockchain integration. While computationally heavier and more complex to deploy, it is particularly well-suited for high-stakes environments such as defense, healthcare, and national identity systems. A hybrid model combining AI for feature extraction and MORSA for secure authentication and logging may offer a balanced, next-generation framework for robust voice authentication.

From this Table 12, Conclude that

➢

In the breath deep scenario, the voice samples that contain only breathing deep sounds.

➢

shallow breath scenario was performed by using voice samples that contain only shallow breathing sounds.

➢

In all breaths, the voice samples used contain both deep breathing and shallow breathing sounds

Table 13. Comparative analysis identity management system utilizing voice authentication.

Table 13. Comparative analysis identity management system utilizing voice authentication.

Hidden neurons	TP	TN	FP	FN	Accuracy	Precision	Recall	Specificity	F-measure	G-mean	Execution time (s)
Deep breath	150	11	12	01	95.83	100.0	91.67	100.00	95.65	95.74	119.8567
Shallow breath	300	10	12	02	91.67	100.0	83.33	100.00	90.91	91.29	119.5190
All breath	250	22	19	41	89.13	84.62	95.65	82.61	89.8	89.96	132.7839

E. Comparative Analysis with Existing Techniques

To assessing the proposed approach's performance, the current voice authentication approaches are taken into consideration, including MFCC or Mel-frequency cepstral coefficients, HMM (Hidden Markov mode) , Gaussian Mixture Model (GMM) [178]. In this part, we can see how different performance metrics were used to compare the designed JHBO driven DNFN with data (training) and k-fold values shown in Table 2 and 3. Predicted from other method, a proposed method using DCNN kernel A DNFN K-fold fusion protein based on JHBO Verifying precision 0.9176, 0.9005, 0.8806, 0.8894. A sensitivity of 0.8982, 0.8816, 0.8938, an alpha of 0.9307 Findings with a high degree of certainty (0.9125, 0.8926, 0.9014, 0.9219) Information used for training Accuracy of testing: 0.8959, 0.8910, 0.8901, 0.9151 In terms of specificity, we have 0.902 0.8896 0.8932 0.9182 and a sensitivity of 0.9123 0.8868 0.9001 0.9218. The performance measures and the designed proposed method using JHBO-based DNFN are compared in Table 2. The results of a comparison study of JHBO-driven DNFN that were introduced for assessing accuracy with different values of k-fold. In terms of k-fold value, MFCC, HMM, and GMM all achieve testing accuracy of 0.8925, with MFCC coming in at 0.887, and the proposed JHBO-based DNFN reaching 0.9054.

A comparison of the sensitivity of JHBO-based DNFN with different k-fold values for k-fold value, MFCC, HMM, and GMM have sensitivity values of 0.8912, 0.878, and 0.8816, respectively; in contrast, created proposed method using JHBO-based DNFN has a sensitivity of 0.9085.

Table 5.4. Comparative Performance with Other Methods.

Table 5.4. Comparative Performance with Other Methods.

Metric	MFCC Based	HMM	GMM	Proposed
Accuracy	0.8956	0.8910	0.8910	0.8979
Sensitivity	0.885	0.8831	0.8813	0.9107
Specificity	0.9164	0.8963	0.9031	0.9221

Table 13. In a k-fold analysis, the proposed method continued to show superior performance.

Table 13. In a k-fold analysis, the proposed method continued to show superior performance.

Metric	MFCC	HMM	GMM	Proposed
Accuracy	0.9176	0.9005	0.8806	0.9394
Sensitivity	0.8982	0.8816	0.8938	0.9307
Specificity	0.9125	0.8926	0.9014	0.9312

V. Conclusion

References

1. Chakraborty M, Mallick, AK.AES Encrypted FSK Generation at X-Band Frequency using a Single Reflex Klystron. Wireless Communication over ZigBee for Automotive Inclination Measurement. China Communications 2010, 7, 1–9.
2. Pal P, Chandra Sahana B, Poray J. RSA encrypted FSK RF transmission powered by an innovative microwave technique for invulnerable security. The Journal of Defense Modeling and Simulation. August 2022. [CrossRef]
3. Pal, P. , Sahana, B.C., Poray, J. MORSA: An innovative modified and optimized RSA framework using parallel computing environment. Mathematical Modelling of Engineering Problems 2025, 12, 2085–2096. [Google Scholar] [CrossRef]
4. Reynolds, “Voice biometrics: Feature extraction methods. IEEE Trans. Biomed. Eng. 2019, 65, 1132–1140.
5. Y. Zhao et al., “Deep learning for speaker recognition. IEEE Signal Process. Lett. 2020, 26, 1551–1555.
6. T. N. Sainath et al., “Convolutional, Long Short-Term Memory, fully connected Deep Neural Networks,” ICASSP, 2015.
7. N. Koblitz, “Cryptographic methods for biometric data protection, Cryptologia 2016, 30, 52–61.
8. P. K. Bhatia et al., “Multi-prime RSA and its performance evaluation. IJCNIS 2019, 3, 23–29.
9. J. Proakis, Digital Communications, 5th ed., McGraw-Hill, 2007.
10. S. Kumar and R. Jain, “FSK modulation for robust audio data transmission. IEEE Comm. Lett. 2020, 24, 540–544.
11. T. Kinnunen and H. Li, “An overview of text-independent speaker recognition: From features to supervectors. Speech Communication 2020, 52, 12–40.
12. L. R. Rabiner and B. H. Juang, Fundamentals of Speech Recognition, Prentice Hall, 1993.
13. Menezes, P. van Oorschot, and S. Vanstone, Handbook of Applied Cryptography, CRC Press, 1996.
14. P. K. Bhatia et al., “Multi-prime RSA and its performance evaluation. IJCNIS 2011, 3, 23–29.
15. M. L. Sisodia and V. L. Gupta, Microwave Engineering, New Age International, 2007.
16. J. G. Proakis, Digital Communications, 5th ed., McGraw-Hill, 2007.
17. Jain et al., “Biometric template security. EURASIP Journal on Advances in Signal Processing 2008, 2008, 1–17.
18. M. Crosby et al., “Blockchain technology: Beyond bitcoin. Applied Innovation Review 2022, 6–10.
19. Dar JA, Srivastava KK, Ahmed Lone S. Design and development of hybrid optimization enabled deep learning model for COVID-19 detection with comparative analysis with DCNN, BIAT-GRU, XGBoost. Comput Biol Med. 2022, 150, 106123. [CrossRef] [PubMed]
20. Yuechi Jiang, Frank H.F. Leung, Investigating and improving the utility of probabilistic linear discriminant analysis for acoustic signal classification. Digital Signal Processing 2021, 114, 103055. [CrossRef]