Aura-CAPTCHA: A Reinforcement Learning and GAN-Enhanced Multi-Modal CAPTCHA System

Joydeep Chandra; Prabal Manhas; Ramanjot Kaur; Rashi Sahay

arxiv: 2508.14976 · v2 · submitted 2025-08-20 · 💻 cs.LG

Aura-CAPTCHA: A Reinforcement Learning and GAN-Enhanced Multi-Modal CAPTCHA System

Joydeep Chandra , Prabal Manhas , Ramanjot Kaur , Rashi Sahay This is my paper

Pith reviewed 2026-05-18 21:41 UTC · model grok-4.3

classification 💻 cs.LG

keywords CAPTCHAGANReinforcement LearningMulti-modal verificationBehavioral analysisBot detectionAdaptive challenges

0 comments

The pith

Aura-CAPTCHA uses GAN-generated stimuli, reinforcement learning for real-time difficulty adjustment, and behavioral analysis to raise human success rates while cutting classical bot bypass rates.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents Aura-CAPTCHA as a multi-modal system that generates unique visual challenges with GANs, pairs them with audio, and lets a reinforcement learning agent change difficulty according to live user interaction data. A hybrid classifier then combines fixed rules with machine learning to label interactions as human or bot. The authors compare the approach to text CAPTCHAs, reCAPTCHA v2, audio schemes, and invisible risk systems, testing it against CNN solvers, YOLO detectors, and vision-language agents. If the results hold, sites could replace static tests with adaptive ones that let more legitimate users pass yet block more automated attacks than before.

Core claim

Aura-CAPTCHA synthesizes fresh visual stimuli through GANs, synchronizes them with audio challenges, and employs an RL agent to tune difficulty from real-time behavioral signals; its hybrid heuristic-plus-machine-learning classifier then separates human from bot sessions, delivering higher human completion rates and lower success for documented convolutional, detection, and agentic attacks than static baselines.

What carries the argument

The RL-driven adaptive challenge engine that alters difficulty on the fly using interaction patterns, together with the GAN-based stimulus generator and the hybrid classifier that fuses rules with learned behavioral features.

If this is right

Websites could replace fixed-image or text CAPTCHAs with ones whose difficulty shifts automatically to each visitor.
Classical deep-learning attacks such as CNN solvers and YOLO pipelines would succeed less often against the generated stimuli.
Real-time behavioral signals would let the system respond to suspicious patterns before a session completes.
The same multi-modal generation plus adaptation pattern could be applied to other verification tasks that currently rely on static tests.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The approach illustrates an ongoing arms race in which explicit challenge systems must keep generating fresh gaps that current large models have not yet closed.
Combining the classifier with existing invisible risk-scoring layers could reduce the number of explicit challenges shown to low-risk users.
Longer-term security may depend on moving beyond perceptual tests toward tasks that exploit cognitive differences between humans and current agents.
The reported gains in human success rate would need re-testing on mobile devices and across age groups to confirm they generalize.

Load-bearing premise

The hybrid classifier can separate genuine human behavior from bot behavior in real time without generating enough false positives to block ordinary users.

What would settle it

A controlled trial that measures the fraction of verified human participants who fail the Aura-CAPTCHA due to the classifier, or that shows bypass rates for the full system equal or exceed those of the static baselines when the same attack models are applied.

Figures

Figures reproduced from arXiv: 2508.14976 by Joydeep Chandra, Prabal Manhas, Ramanjot Kaur, Rashi Sahay.

**Figure 1.** Figure 1: Methodology of Aura-CAPTCHA 3.1 System Architecture Aura-CAPTCHA comprises three core modules: the Generative Content Module, the Adaptive Challenge Module, and the User Interaction Analysis Module. These modules collaborate to generate dynamic, context-aware challenges, adapt difficulty in real-time, and analyze user interactions for enhanced bot detection. In [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Figure 2: Architecture of Aura-CAPTCHA [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Outputs of Aura Captcha System generated Captchas 4.2 Mathematical Model of Aura-CAPTCHA The core of Aura-CAPTCHA is built upon a Reinforcement Learning (RL) model that dynamically adjusts the difficulty level of the CAPTCHA challenge based on user interaction data. The Q-learning approach is defined as: Q(s, a) ← Q(s, a) + α h r + γ max a′ Q(s ′ , a′ ) − Q(s, a) i (1) where: – Q(s, a): The Q-value represe… view at source ↗

**Figure 4.** Figure 4: Comparison of CAPTCHA Systems Across Key Metrics [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

read the original abstract

We present Aura-CAPTCHA, a multi-modal verification system that integrates Generative Adversarial Networks (GANs), Reinforcement Learning (RL), and behavioral analysis to create adaptive challenges resistant to classical deep-learning attacks. Our system synthesizes unique visual stimuli via GAN-based generation alongside synchronized audio challenges, while an RL agent adjusts difficulty based on real-time user interaction patterns. A hybrid classifier combining heuristic rules and machine learning distinguishes human from bot interactions. We position Aura-CAPTCHA relative to well-established baselines (text-based schemes, Google reCAPTCHA v2, audio alternatives, and modern invisible risk-analysis systems) and evaluate it against documented state-of-the-art attacks, including convolutional-neural-network solvers, object-detection pipelines (YOLO), and recent agentic vision-language models. Experimental results indicate that Aura-CAPTCHA improves human success rates and lowers classical bypass rates compared to static challenge-based baselines, although, like all explicit-challenge systems, it remains vulnerable to emerging large-model agents. We discuss these limitations transparently and outline future directions toward cognitive-gap-based defenses.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Aura-CAPTCHA assembles GAN image generation, RL difficulty tuning, and a hybrid behavioral classifier into one system, but the experimental claims rest on unshown numbers and an untested assumption about low false positives.

read the letter

The core of this paper is a practical integration: GANs create fresh visual challenges, an RL agent tweaks difficulty from live interaction data, and a rules-plus-ML layer tries to separate humans from bots. They also add synchronized audio. That combination is the main thing on offer, and it is presented clearly against text CAPTCHAs, reCAPTCHA v2, and attacks such as CNN solvers or YOLO detectors. The authors note that large vision-language models can still break through, which keeps the discussion grounded rather than overstated. The multi-modal setup and the adaptive RL component are the parts that feel like they could be useful to someone actually shipping a CAPTCHA service. They give a reasonable sense of how the pieces fit together without claiming a new theoretical foundation. The citation pattern looks standard for this area and does not appear to skip key prior work on GAN-based challenges or behavioral signals. The soft spot is the evaluation. The abstract states better human success rates and lower bypass rates, yet supplies no concrete figures, dataset sizes, or ablation results. The hybrid classifier is described but not measured for false-positive behavior on varied human sessions or devices. If the full paper contains those metrics and shows they hold up, the central claim strengthens; without them the reported gains could be an artifact of how the detector was tuned. This is the kind of work that belongs in a reading group focused on applied security tools rather than core machine-learning theory. A practitioner looking for implementation ideas on adaptive challenges would get value from the architecture description. A serious editor should send it to peer review so the authors can supply the missing quantitative checks on the classifier and the human trials. The system is concrete enough to be worth referee time even if the current evidence is thin.

Referee Report

2 major / 2 minor

Summary. The paper introduces Aura-CAPTCHA, a multi-modal CAPTCHA that uses GANs to synthesize adaptive visual stimuli, an RL agent to adjust challenge difficulty from real-time interaction patterns, synchronized audio challenges, and a hybrid heuristic-plus-ML classifier to separate human from bot behavior. It positions the system against text CAPTCHAs, reCAPTCHA v2, audio schemes, and invisible risk-analysis systems, and reports experimental comparisons against CNN solvers, YOLO pipelines, and agentic vision-language models, claiming higher human success rates and lower classical bypass rates while acknowledging vulnerability to large-model agents.

Significance. If the experimental claims are substantiated with quantitative metrics, the work could offer a concrete step toward generative and behaviorally adaptive defenses that raise the bar for automated attacks beyond static challenges. The explicit integration of GAN generation, RL adaptation, and multi-modal (visual-audio) cues is a clear engineering contribution, though its impact hinges on whether the hybrid classifier demonstrably avoids excessive false positives on legitimate users.

major comments (2)

[Hybrid Classifier] Hybrid classifier subsection: the central claim that Aura-CAPTCHA improves human success rates relative to baselines rests on the classifier correctly labeling legitimate users without high false-positive rates that would trigger re-challenges. No FPR, precision-recall, or ablation results on held-out human sessions (across devices or demographics) are supplied, so any reported success-rate gain could be an artifact of detector permissiveness rather than the GAN/RL design.
[Experimental Evaluation] Experimental evaluation section: the abstract states performance gains against CNN solvers, YOLO, and agentic vision-language models, yet supplies no success rates, error bars, dataset sizes, attack exclusion criteria, or statistical tests. Without these numbers the comparative claims cannot be verified or reproduced.

minor comments (2)

[RL Agent] Notation for the RL state and reward functions is introduced without an explicit equation or pseudocode block, making the adaptation mechanism harder to follow.
[Figures] Figure captions for the GAN-generated stimuli and audio waveforms should include the exact generation parameters and synchronization method used in the experiments.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. The comments identify key areas where additional quantitative detail will strengthen the presentation of the hybrid classifier and experimental results. We respond to each major comment below and indicate the revisions we will make.

read point-by-point responses

Referee: [Hybrid Classifier] Hybrid classifier subsection: the central claim that Aura-CAPTCHA improves human success rates relative to baselines rests on the classifier correctly labeling legitimate users without high false-positive rates that would trigger re-challenges. No FPR, precision-recall, or ablation results on held-out human sessions (across devices or demographics) are supplied, so any reported success-rate gain could be an artifact of detector permissiveness rather than the GAN/RL design.

Authors: We agree that explicit performance metrics for the hybrid classifier are required to support the reported human success rates. The manuscript currently emphasizes end-to-end system results rather than isolated classifier diagnostics. In the revised version we will add FPR, precision-recall curves, and an ablation isolating the classifier's contribution, computed on held-out human sessions. Device-level breakdowns will be reported from the data we collected; demographic breakdowns were not recorded in the original study and will be noted as a limitation. revision: partial
Referee: [Experimental Evaluation] Experimental evaluation section: the abstract states performance gains against CNN solvers, YOLO, and agentic vision-language models, yet supplies no success rates, error bars, dataset sizes, attack exclusion criteria, or statistical tests. Without these numbers the comparative claims cannot be verified or reproduced.

Authors: We acknowledge that the experimental section would be more verifiable with fuller numerical reporting. The manuscript contains comparative tables, but we will expand the text to include per-attack success rates with error bars, exact dataset sizes (human sessions and attack attempts), explicit attack exclusion criteria, and results of statistical tests (e.g., paired t-tests). These additions will be placed in the revised experimental evaluation section. revision: yes

Circularity Check

0 steps flagged

No circularity; experimental claims do not reduce to inputs by construction.

full rationale

The paper describes a multi-modal CAPTCHA system using GANs for stimulus generation, RL for difficulty adjustment, and a hybrid classifier for bot detection. Central claims rest on experimental comparisons of human success rates and bypass rates against baselines and attacks, with no equations, fitted parameters, or derivations presented that would reduce predictions to inputs by construction. No self-citations, ansatzes, or uniqueness theorems are invoked in a load-bearing way within the abstract or positioning. The work is self-contained as an engineering and evaluation contribution rather than a mathematical derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The central claim implicitly assumes that behavioral signals provide an independent signal separable from the visual and audio challenges.

pith-pipeline@v0.9.0 · 5724 in / 1120 out tokens · 52575 ms · 2026-05-18T21:41:45.842198+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

The core of Aura-CAPTCHA is built upon a Reinforcement Learning (RL) model that dynamically adjusts the difficulty level... Q(s,a) update and reward r = +1/-1/0 based on response correctness and timing.
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean LogicNat recovery unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

hybrid classifier combining heuristic rules and machine learning... SVM decision function f(x) = w·x + b on features [avg_time_interval, std_time_interval, ...]

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages

[1]

International Journal of Security and Networks, vol

Hossen, A., Hei, X.: Deepfake CAPTCHA: A Method for Preventing Fake Calls. International Journal of Security and Networks, vol. 16, no. 3, pp. 145–155, 2023

work page 2023
[2]

Chen, Y., Zhang, L., Wang, X.: High-Quality Visually-Guided Sound Separation from Diverse Categories.Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023

work page 2023
[3]

Patel, R., Singh, T., Verma, A.: Improving the Security of Audio CAPTCHAs With Adversarial Examples.Journal of Machine Learning Research, vol. 24, pp. 112–130, 2023

work page 2023
[4]

Huang, J., Zhao, W., Li, K.: Audiobox: Unified Audio Generation with Natural Language Prompts.Neural Networks, vol. 154, pp. 132–145, 2023

work page 2023
[5]

Ars Technica, Sep

Jones, A., White, D.: AI bots now beat 100% of those traffic-image CAPTCHAs. Ars Technica, Sep. 2024. Available at: https://arstechnica.com/ai/2024/09/ ai-defeats-traffic-image-captcha/

work page 2024
[6]

Multimedia Tools and Applications, vol

Kumar, P., Singh, S.: Generative Artificial Intelligence: A Systematic Review and Applications. Multimedia Tools and Applications, vol. 81, no. 14, pp. 19571–19599, 2024

work page 2024
[7]

Proceedings of the 17th International Conference on Artificial Intelligence, pp

Lee, C., Kim, H.: Multi-language Audio-Visual Content Generation Based on GANs. Proceedings of the 17th International Conference on Artificial Intelligence, pp. 77–86, 2023. 12 Authors Suppressed Due to Excessive Length

work page 2023
[8]

Hossen, A., Hei, X.: aaeCAPTCHA: Audio Adversarial CAPTCHA Design.IEEE Transactions on Information Forensics and Security, vol. 19, pp. 788–798, 2024

work page 2024
[9]

Hossen, A., Hei, X.: Adversarial Noise in Audio CAPTCHAs.Springer Journal of Security Research, vol. 12, no. 2, pp. 157–170, 2023

work page 2023
[10]

Yasur, L., Tan, D.: User Behavior Analysis in CAPTCHA Systems.International Journal of Human-Computer Interaction, vol. 38, no. 5, pp. 399–412, 2024

work page 2024
[11]

245–258, 2023

Huang, J., Zhao, W., Li, K.: Audio-Visual CAPTCHA with GANs.Proceedings of the 30th ACM Multimedia Conference, pp. 245–258, 2023

work page 2023
[12]

ACM Computing Surveys, vol

Patel, S., Desai, R.: GAN-Based CAPTCHA Systems: A Comprehensive Review. ACM Computing Surveys, vol. 56, no. 1, pp. 1–23, 2024

work page 2024
[13]

Proceedings of the 12th International Conference on Human-Computer Interaction (HCI), pp

Yasur, L., Tan, D., Ling, H.: Adaptive CAPTCHA with Reinforcement Learning. Proceedings of the 12th International Conference on Human-Computer Interaction (HCI), pp. 300–315, 2024

work page 2024
[14]

Chen, K., Liu, Y.: Dynamic Difficulty Adjustment in CAPTCHA Systems.IEEE Transactions on Cybernetics, vol. 55, no. 8, pp. 895–908, 2024

work page 2024
[15]

112–121, 2023

Smith, A., Jones, M.: SVM Classifier for User Interaction Analysis.Proceedings of the 18th International Conference on Pattern Recognition (ICPR), pp. 112–121, 2023

work page 2023
[16]

Yasur, L., Ling, H., Tan, D.: Real-Time CAPTCHA Adaptation.Journal of Arti- ficial Intelligence Research, vol. 62, no. 1, pp. 178–191, 2024

work page 2024
[17]

IEEE Transactions on Neural Networks and Learning Systems, vol

Zhao, Y., Wang, X., Li, K.: AudioGPT: Generative Model for Synchronized Audio. IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 7, pp. 987–1001, 2024

work page 2024
[18]

45–56, 2023

Huang, J., Zhao, W.: Audiobox: Multi-Modal Content Generation.Proceedings of the European Conference on Artificial Intelligence (ECAI), pp. 45–56, 2023

work page 2023
[19]

Zhao, Y., Li, K., Chen, H.: Visual Echoes: Unified Transformer Model for Syn- chronization.Neural Information Processing Systems (NeurIPS), vol. 36, pp. 1500– 1512, 2024

work page 2024
[20]

Available at: https://openai.com/research/gpt-whisper, 2024

OpenAI: GPT and Whisper: Advanced AI Models in CAPTCHA Solving. Available at: https://openai.com/research/gpt-whisper, 2024

work page 2024
[21]

Liu, C., Zhang, T.: Accessibility in Multi-Modal CAPTCHA Systems.Universal Access in the Information Society, vol. 24, no. 2, pp. 245–257, 2024

work page 2024
[22]

Smith, B., Chang, L.: StyleGAN for CAPTCHA Image Generation.Pattern Recog- nition Letters, vol. 188, pp. 32–45, 2023

work page 2023
[23]

Patel, V., Kumar, S.: Reinforcement Learning in CAPTCHA Design.IEEE Trans- actions on Systems, Man, and Cybernetics, vol. 54, no. 6, pp. 598–612, 2024

work page 2024
[24]

Lee, H., Park, J.: Adaptive Challenge Mechanisms in User Interfaces.Journal of Human-Computer Studies, vol. 152, pp. 301–315, 2023

work page 2023
[25]

Jones, A., Smith, T.: Multi-Modal Synchronization Techniques.IEEE Access, vol. 11, pp. 4567–4581, 2024

work page 2024

[1] [1]

International Journal of Security and Networks, vol

Hossen, A., Hei, X.: Deepfake CAPTCHA: A Method for Preventing Fake Calls. International Journal of Security and Networks, vol. 16, no. 3, pp. 145–155, 2023

work page 2023

[2] [2]

Chen, Y., Zhang, L., Wang, X.: High-Quality Visually-Guided Sound Separation from Diverse Categories.Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023

work page 2023

[3] [3]

Patel, R., Singh, T., Verma, A.: Improving the Security of Audio CAPTCHAs With Adversarial Examples.Journal of Machine Learning Research, vol. 24, pp. 112–130, 2023

work page 2023

[4] [4]

Huang, J., Zhao, W., Li, K.: Audiobox: Unified Audio Generation with Natural Language Prompts.Neural Networks, vol. 154, pp. 132–145, 2023

work page 2023

[5] [5]

Ars Technica, Sep

Jones, A., White, D.: AI bots now beat 100% of those traffic-image CAPTCHAs. Ars Technica, Sep. 2024. Available at: https://arstechnica.com/ai/2024/09/ ai-defeats-traffic-image-captcha/

work page 2024

[6] [6]

Multimedia Tools and Applications, vol

Kumar, P., Singh, S.: Generative Artificial Intelligence: A Systematic Review and Applications. Multimedia Tools and Applications, vol. 81, no. 14, pp. 19571–19599, 2024

work page 2024

[7] [7]

Proceedings of the 17th International Conference on Artificial Intelligence, pp

Lee, C., Kim, H.: Multi-language Audio-Visual Content Generation Based on GANs. Proceedings of the 17th International Conference on Artificial Intelligence, pp. 77–86, 2023. 12 Authors Suppressed Due to Excessive Length

work page 2023

[8] [8]

Hossen, A., Hei, X.: aaeCAPTCHA: Audio Adversarial CAPTCHA Design.IEEE Transactions on Information Forensics and Security, vol. 19, pp. 788–798, 2024

work page 2024

[9] [9]

Hossen, A., Hei, X.: Adversarial Noise in Audio CAPTCHAs.Springer Journal of Security Research, vol. 12, no. 2, pp. 157–170, 2023

work page 2023

[10] [10]

Yasur, L., Tan, D.: User Behavior Analysis in CAPTCHA Systems.International Journal of Human-Computer Interaction, vol. 38, no. 5, pp. 399–412, 2024

work page 2024

[11] [11]

245–258, 2023

Huang, J., Zhao, W., Li, K.: Audio-Visual CAPTCHA with GANs.Proceedings of the 30th ACM Multimedia Conference, pp. 245–258, 2023

work page 2023

[12] [12]

ACM Computing Surveys, vol

Patel, S., Desai, R.: GAN-Based CAPTCHA Systems: A Comprehensive Review. ACM Computing Surveys, vol. 56, no. 1, pp. 1–23, 2024

work page 2024

[13] [13]

Proceedings of the 12th International Conference on Human-Computer Interaction (HCI), pp

Yasur, L., Tan, D., Ling, H.: Adaptive CAPTCHA with Reinforcement Learning. Proceedings of the 12th International Conference on Human-Computer Interaction (HCI), pp. 300–315, 2024

work page 2024

[14] [14]

Chen, K., Liu, Y.: Dynamic Difficulty Adjustment in CAPTCHA Systems.IEEE Transactions on Cybernetics, vol. 55, no. 8, pp. 895–908, 2024

work page 2024

[15] [15]

112–121, 2023

Smith, A., Jones, M.: SVM Classifier for User Interaction Analysis.Proceedings of the 18th International Conference on Pattern Recognition (ICPR), pp. 112–121, 2023

work page 2023

[16] [16]

Yasur, L., Ling, H., Tan, D.: Real-Time CAPTCHA Adaptation.Journal of Arti- ficial Intelligence Research, vol. 62, no. 1, pp. 178–191, 2024

work page 2024

[17] [17]

IEEE Transactions on Neural Networks and Learning Systems, vol

Zhao, Y., Wang, X., Li, K.: AudioGPT: Generative Model for Synchronized Audio. IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 7, pp. 987–1001, 2024

work page 2024

[18] [18]

45–56, 2023

Huang, J., Zhao, W.: Audiobox: Multi-Modal Content Generation.Proceedings of the European Conference on Artificial Intelligence (ECAI), pp. 45–56, 2023

work page 2023

[19] [19]

Zhao, Y., Li, K., Chen, H.: Visual Echoes: Unified Transformer Model for Syn- chronization.Neural Information Processing Systems (NeurIPS), vol. 36, pp. 1500– 1512, 2024

work page 2024

[20] [20]

Available at: https://openai.com/research/gpt-whisper, 2024

OpenAI: GPT and Whisper: Advanced AI Models in CAPTCHA Solving. Available at: https://openai.com/research/gpt-whisper, 2024

work page 2024

[21] [21]

Liu, C., Zhang, T.: Accessibility in Multi-Modal CAPTCHA Systems.Universal Access in the Information Society, vol. 24, no. 2, pp. 245–257, 2024

work page 2024

[22] [22]

Smith, B., Chang, L.: StyleGAN for CAPTCHA Image Generation.Pattern Recog- nition Letters, vol. 188, pp. 32–45, 2023

work page 2023

[23] [23]

Patel, V., Kumar, S.: Reinforcement Learning in CAPTCHA Design.IEEE Trans- actions on Systems, Man, and Cybernetics, vol. 54, no. 6, pp. 598–612, 2024

work page 2024

[24] [24]

Lee, H., Park, J.: Adaptive Challenge Mechanisms in User Interfaces.Journal of Human-Computer Studies, vol. 152, pp. 301–315, 2023

work page 2023

[25] [25]

Jones, A., Smith, T.: Multi-Modal Synchronization Techniques.IEEE Access, vol. 11, pp. 4567–4581, 2024

work page 2024