A novel deep learning-based statistical randomness evaluation test methodology for cryptographic applications
Open accessThe DOI matches its Crossref record; 77 of the 77 references it cites were found in a registry or an index; 5 of the 18 sources its claims cite were checked for support.
Abstract
The security of cryptographic systems is directly linked to the statistical randomness properties of the random numbers used. Traditional statistical randomness tests can be limited in evaluating the properties of these numbers and can require long processing times. While widely used test suites such as the National Institute of Standards and Technology (NIST) Special Publication (SP) 800–22 play a crucial role in assessing data randomness, they are slow on large data sets, can only evaluate certain statistical properties, and fail to detect complex data patterns and dependency structures. In this paper, we propose a new deep learning (DL) based method to overcome these limitations. In the study, bit sequences produced by an FPGA-based true random number generator (TRNG) and different pseudo-random number generators (PRNGs) were converted into image format, and classification experiments were carried out on the AlexNet, ResNet50, and EfficientNetB0 architectures. The results showed that AlexNet, with $$87\boldsymbol{\%}$$ accuracy and $$99\boldsymbol{\%}$$ recall, was by far the most successful method compared to the other models. Furthermore, the ablation analysis revealed that the components of data augmentation, early stopping, and cross-validation play a critical role in the model’s stability and generalizability. Providing a reliable randomness evaluation with high accuracy, this new method, as an alternative to traditional statistical test suites, demonstrates that AI-based solutions can enhance the effectiveness and accuracy of randomness assessment in cryptographic applications. The success of DL methods in complex data structures offers significant potential for improving the security of modern cryptographic systems.
Overview
CoChat Summary
TL;DR
The paper proposes a deep learning-based statistical randomness evaluation method that converts bit sequences from FPGA-based TRNG and PRNG sources into 224 × 224 RGB images and classifies them with CNN architectures. AlexNet achieved the best performance, with 87% accuracy and 99% recall, outperforming ResNet50 and EfficientNetB0.
Key findings
- AlexNet achieved 87% accuracy, 79% precision, 99% recall, and 88% F1 score; its TPR was 0.99, FPR 0.25, and AUC 0.87.
- ResNet50 performed near random guessing: 50% accuracy, 50% precision, 100% recall, 67% F1 score, TPR 1.00, FPR 1.00, and AUC 0.50. EfficientNetB0 reached 56% accuracy, 57% precision, 89% recall, 69% F1 score, TPR 0.89, FPR 0.77, and AUC 0.56.
- SVM and Random Forest reached 84% overall accuracy but had low precision (76–77%) for non-secure sequences failing NIST tests; AlexNet exceeded 87% accuracy with a more balanced precision-recall relationship and higher F1 scores.
- Ablation analysis on AlexNet showed removing data augmentation increased average validation accuracy to 94.72% with a 5.52% standard deviation; removing early stopping increased validation accuracy only marginally to 88.16%; omitting cross-validation resulted in 100% accuracy, suggesting potential overlap between training and test data or insufficient test data for proper generalization.
- Complexity measures: raw FPGA data had SE = 13.52, C0 = 0.0017, PE = 0.7705; PRNG sequences failing randomness tests had lower complexity (SE = 12.00, PE = 0.5977); post-processed FPGA sequences had SE = 13.57, PE = 0.7737; PRNG sequences passing tests had SE = 13.57, PE = 0.7736.
Methods
- Dataset: bit sequences from an FPGA-based RO-TRNG (Altera Cyclone IV EP4CGX150DF31), including raw and post-processed FPGA data, plus PRNG sequences passing NIST SP 800-22 tests (generated with Python’s secrets library) and PRNG sequences failing NIST tests (local bias method, bias = 0.8). The balanced dataset contained 2,000 sample files per class, totaling 8,000 files; 6,400 images were used for training and 1,600 for validation, with each class containing 3,200 training and 800 validation images. Each sequence of 1,204,224 bits was converted into a 224 × 224 RGB image using 24-bit pixels (8 bits per channel).
- Models and training: AlexNet (17-layer optimized version), ResNet50, and EfficientNetB0 were compared. Hyperparameters included learning rate 0.001 updated every 5 epochs, 100 epochs, batch size 16, Adam optimization, 5-fold cross-validation, early stopping after 10 epochs without improvement, 50% dropout, and data augmentation by rotations of 90°, 180°, and 270°. Training used an 80:20 training/validation split on an 11th Gen Intel Core i7-11800H @ 2.30 GHz system.
- Analysis: NIST SP 800-22 was used for labeling sequences as passed or not passed. Model evaluation used accuracy, precision, recall, F1 score, TPR, FPR, and AUC. Bootstrap analysis used 1,000 iterations with 80% of the test set per sample. Spectral entropy (SE), C0 complexity, and permutation entropy (PE) were also calculated.
Limitations
- The paper does not present a dedicated limitations section. It states as future work that the methodology can be extended to cryptanalysis-oriented scenarios such as adversarial robustness, side-channel analysis, and known-plaintext attacks, planned for subsequent studies.
- The text makes evident that EfficientNetB0 and ResNet50 overfit the limited dataset, resulting in lower validation performance; ResNet50 performed close to random guessing.
- The ablation analysis indicates that omitting cross-validation produced 100% accuracy, which the authors interpret as suggesting potential overlap between training and test data or insufficient test data for proper generalization. Removing data augmentation increased average validation accuracy but with a high standard deviation of 5.52%, implying reduced stability.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
1 Valid · 12 Not valid · 11 Not checked · 140 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“In particular, the ability to detect complex dependency structures and identify long-term data relationships makes the integration of CNNs into cryptographic security systems important (Priya et al. 2024).”
§ A novel deep learning-based statistical randomness evaluation test methodology for cryptographic applications > 2.3 Deep learning and convolutional neural network architecture
What the sources say
- Partly supportedcites Priya et al. 2024 sourceExperimental — accuracy not yet measured
The State of art section says incorporating neural networks can make image encryption more resilient to attacks, and the Proposed methodology describes convolutional layers within a system that also uses cryptography (page numbers unavailable). Neither passage attributes that benefit to CNNs detecting complex dependency structures or identifying long-term data relationships. (LLM, conf=0.87)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“In particular, DL algorithms can perform fast and highly accurate evaluations on large datasets thanks to their ability to detect complex dependency structures and unpredictable patterns”
§ A novel deep learning-based statistical randomness evaluation test methodology for cryptographic applications > 1 Introductıon
What the sources say
- Could not verifycites Coutinho et al. 2018 sourceExperimental — accuracy not yet measured
The Introduction says AI systems have excelled at tasks such as image and speech recognition, but does not describe fast, accurate evaluation of large datasets or attribute it to detecting complex dependencies and unpredictable patterns. The Results (§4.1; page unavailable) instead describes training as time-consuming for large keys. Searches for the claim’s distinctive terms and paraphrases found no passage establishing the attributed explanation. (LLM, conf=0.73)
- Not checked — no full text of the sourcecites Khalid et al. 2023 source
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“The use of CNN in randomness testing offers significant advantages in overcoming the shortcomings of traditional statistical tests.”
§ A novel deep learning-based statistical randomness evaluation test methodology for cryptographic applications > 2.3 Deep learning and convolutional neural network architecture
What the sources say
- Could not verifycites Priya et al. 2024 sourceExperimental — accuracy not yet measured
The NIST-SP 800-22 analysis section uses statistical tests to assess cipher-image randomness, while the Statistical analysis section describes entropy and other image-security measures. Searches for CNN-based randomness testing and comparisons with traditional statistical tests found no passage addressing the claimed advantage. (LLM, conf=0.76)
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The findings of the study demonstrate that deep learning methods yield effective results across critical performance metrics such as accuracy, precision, sensitivity, and processing speed.”
§ A novel deep learning-based statistical randomness evaluation test methodology for cryptographic applications > 1 Introductıon
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Furthermore, the ablation analysis revealed that the components of data augmentation, early stopping, and cross-validation play a critical role in the model's stability and generalizability.”
§ A novel deep learning-based statistical randomness evaluation test methodology for cryptographic applications > Abstract
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
62 more claims
Create a free account to see all 164 claims and to verify the 1 reference still open.
- It may state two separate claims as one. 69
- The quote only partly states this claim. 3
- Probably not a substantial claim. 46
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ A novel deep learning-based statistical randomness evaluation test methodology for cryptographic applications > 4.1 Statistical reliability analysis of model performances
“the F1 scores for the EfficientNetB0 model are clustered around 0 , demonstrating that this model failed to perform effective classification on the given dataset.”
§ A novel deep learning-based statistical randomness evaluation test methodology for cryptographic applications > 5 Results and conclusıons
“| EfficientNetB0 | 0.56 | 0.57 | 0.89 | 0.69 |”
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 1×
“In this study, different deep learning architectures were comparatively evaluated to enhance randomness testing, specifically employing the AlexNet (Krizhevsky et al. 2017), ResNet50 (He et al. 2016) and EfficientNetB0 (Tan and Le 2019) models.”
ImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155 · cited 1×
“In this study, different deep learning architectures were comparatively evaluated to enhance randomness testing, specifically employing the AlexNet (Krizhevsky et al. 2017), ResNet50 (He et al. 2016) and EfficientNetB0 (Tan and Le 2019) models.”
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
Tan, M., & Le, Q. V. (2019) · arXiv · doi:10.48550/arxiv.1905.11946 · Cited by 17,072
A fast learning algorithm for deep belief nets
Hinton, G. E., Osindero, S., & Teh, Y.-W. (2006) · Neural Computation · doi:10.1162/neco.2006.18.7.1527 · Cited by 16,336 · cited 1×
“DL architectures such as AlexNet, with their automatic learning processes and feedback mechanisms, both improve evaluation accuracy and provide significant advantages in cryptographic applications (Hinton et al. 2006).”
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2015) · arXiv · doi:10.48550/arxiv.1512.03385 · Cited by 4,196
Bassham, L. E., Rukhin, A. L., Soto, J., Nechvatal, J. R., Smid, M. E., Barker, E. B., Leigh, S. D., Levenson, M., Vangel, M., Banks, D. L., et al. (2010) · NIST Special Publication 800-22 Revision 1a · doi:10.6028/nist.sp.800-22r1a · Cited by 3,617 · cited 3×
“…the reliability and statistical adequacy of random number generators, many test methods have been developed, including Diehard (Marsaglia 2003), NIST SP 800–22 (Bassham et al. 2010), AIS20/AIS31 (Schindler 2023), ENT (Walker 2008), FIPS 140–2 (Caddy 2011) and TestU01 (L’Ecuyer and Simard 2007).”
TestU01: A C library for empirical testing of random number generators
L'Ecuyer, P., & Simard, R. (2007) · ACM Transactions on Mathematical Software · doi:10.1145/1268776.1268777 · Cited by 1,014 · cited 6×
“However, these traditional methods have been reported to be inadequate, especially for high-dimensional data and for analyzing complex dependencies (L’Ecuyer and Simard 2007).”
InfluentialA Provably Secure True Random Number Generator with Built-In Tolerance to Active Attacks
Sunar, B., Martin, W. J., & Stinson, D. R. (2007) · IEEE Transactions on Computers · doi:10.1109/tc.2007.250627 · Cited by 673
Privacy-preserving artificial intelligence in healthcare: Techniques and applications
Khalid, N., Qayyum, A., Bilal, M., Al-Fuqaha, A., & Qadir, J. (2023) · Computers in Biology and Medicine · doi:10.1016/j.compbiomed.2023.106848 · Cited by 581 · cited 1×
“In particular, DL algorithms can perform fast and highly accurate evaluations on large datasets thanks to their ability to detect complex dependency structures and unpredictable patterns (Coutinho et al. 2018; Khalid et al. 2023).”
Recommendation for the Entropy Sources Used for Random Bit Generation
Turan, M. S., Barker, E., Kelsey, J., McKay, K. A., Baish, M. L., & Boyle, M. (2018) · NIST Special Publication 800-90B · doi:10.6028/nist.sp.800-90b · Cited by 489 · cited 1×
“In addition, NIST SP 800-90B (Turan et al. 2018) provides minimum entropy estimation methods to evaluate the quality of entropy sources.”
Research works printed but not linked: 4
- Kishore RH, Sekhar AC, Patro P, Chaganti P (2023) A novel homomorphic and matrix operation for randomization encryption schemes for privacy in cloud computing architecture. J Theor Appl Inf Technol 101:1038–1053
- Knuth DE (1997) The art of computer programming: fundamental algorithms, volume 2. Addison-Wesley Professional, Boston, MA
- Schindler W (2023) Overview of AIS 20/31. Presentation at NIST Workshop on Random Number Generation, Gaithersburg, MD. Available at: https://csrc.nist.gov/csrc/media/Presentations/2023/overview-of-ais-2031/images-media/session-2-schindleroverview-of-ais-20-31.pdf
- Wu Y, Noonan JP, Agaian S (2011) NPCR and UACI randomness tests for image encryption. Cyber Journals: Multidisciplinary Journals in Science and Technology, Journal of Selected Areas in Telecommunications (JSAT) 1(2):31–38. Available at: https://www.cyberjournals.com/Papers/Apr2011/05.pdf
Cited by
Hashim, H. R. (2026) · Computation · doi:10.3390/computation14080175
“Note that for more details about the sources of the used metrics in all of the following sections, we refer to [20–22].”
A Unique Technique to RGB Image Cryptography via Design of Deep Learning
Khayrullayev, I., Ashirova, A. I., Bag, S., Rakshit, P., Yuldoshev, J., Sаpаyev, V., & Rajput, B. J. (2026) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-3693-3751-6.ch005
Goswami, A., Singh, R. B., Mishra, S. G., Rishi, K., & Gupta, M. D. (2026) · Integration · doi:10.1016/j.vlsi.2026.102832 · Cited by 1
Xiang, S., Yao, L., Zhang, B., Wu, H., Pan, R., Zhang, R., Xing, Z., Gong, M., & Xu, Y. (2026) · Integration · doi:10.1016/j.vlsi.2026.102808
Towards Automated Identification of Block Cipher Structures Using Machine Learning
Kiran, U., Butt, H. T., & Jalil, Z. (2026) · Journal of Information assurance and security · doi:10.2478/ias-2026-0007
“Kaner et al. [31] used deep learning for randomness analysis instead of relying on NIST randomness with 87 precision and 97 recall and showed another more efficient way of randomness evaluation.”
Metadata
- Authors
- Kaner, S., Garipcan, A. M., & Erdem, E.
- Affiliations
- Muş Alparslan University (TR)
- Fırat University (TR)
- Corresponding author
- Erdem, Ebubekir
- Venue
- Journal of King Saud University Computer and Information Sciences
- Published
- 2025-10-17
- Volume
- 37
- Issue
- 9
- ISSN
- 1319-1578 (ISSN-L)
- 2213-1248
- OpenAlex
- W4415282237
- Cited by
- 11
- References
- 68
- Citations by year
- 2026: 11
- FWCI
- 4.41
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by-nc-nd
- creativecommons.org/licenses/by-nc-nd/4.0 (tdm)
- creativecommons.org/licenses/by-nc-nd/4.0 (vor)
- Locations
- Journal of King Saud University - Computer and Information Sciences (publishedVersion, cc-by-nc-nd)
- DSpace@MŞU (submittedVersion, cc-by-nc-nd)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- DSpace - FIRAT (Fırat University) (submittedVersion, cc-by-nc-nd)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Cryptographic Implementations and Security
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Chaos-based Image/Signal Encryption
- Physical Sciences › Computer Science › Hardware and Architecture › Physical Unclonable Functions (PUFs) and Hardware Security
- Keywords
- statistical randomness testing, random number generators, true random number generator, pseudorandom number generators, image classification, AlexNet, NIST SP 800-22, cryptographic security, data augmentation
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-09-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-09-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: A novel deep learning-based statistical randomness evaluation test methodology for cryptographic applications. https://app.cochat.ai/papers/10.1007/s44443-025-00271-4
- The paper
- Kaner, S., Garipcan, A. M., & Erdem, E. (2025). A novel deep learning-based statistical randomness evaluation test methodology for cryptographic applications. Journal of King Saud University Computer and Information Sciences. https://doi.org/10.1007/s44443-025-00271-4
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.