Home / Current Issue / Paper 1706618
A Hybrid Neuro-Symbolic Framework for Real-Time Detection of Adversarial Attacks in Autonomous Systems
Subject area: Science,Engineering and Technology · Area of research: Hybrid Neuro-Symbolic Framework
Abstract
Rising dependence on autonomous systems in sensitive sectors like self-driving vehicles, smart surveillance, and industrial automation has revealed weaknesses to hostile attacks. Though accurate, traditional deep learning systems still run the risk of adversarial perturbations that might lead artificial intelligence driven decision making off track. To improve adversarial attack detection in real time systems, this paper offers a hybrid neurosymbolic framework that combines the logical reasoning strengths of symbolic AI with the pattern recognition abilities of neural networks. To address hazards in independent cyber physical environments, the suggested approach combines neural feature extraction, symbolic logic based validation, and dynamic adversarial mitigation techniques for threat identification and response. Compared to conventional deep learning models, our approach is judged by means of standard adversarial attack benchmarks, which show greater sensitivity, superior interpretability, and greater resistance against adversarial techniques. This study emphasizes the capacity of neurosymbolic AI to protect autonomous systems and reduce adversarial dangers in mission critical uses.
Keywords
Neuro-symbolic AI, adversarial attacks, autonomous systems, symbolic reasoning, neural networks, adversarial defense, real-time detection, deep learning security.
References
[1] Summary
[2] This section epitomized the proposed hybrid neuro-symbolic technique for detecting adversarial attacks, primarily due to the synergistic effects of symbolic reasoning and deep learning. Additionally, the method has been implemented with some sophistication concerning data pre- filtering, feature extraction, attack classification, and real-time decision-making. Combining feature extraction with neural networks and structured decision-making with symbolic reasoning would result in a complete strategy for adversary domain detection. The framework would, in turn, be suited for use in improving the detection accuracy and interpretability of crucial autonomous systems. The experimental implementation and testing of this proposed model are further illustrated in the next paragraph.
[3] EXPERIMENTAL SETUP AND RESULTS
[4] Experimental Environment
[5] A hybrid neuro-symbolic architecture was employed in order to be tested under controlled experiment conditions, evaluating the system's role in autonomous and real-time adverse system attack detection. The system was coded completely with Python programming language, where the neural network parts are coded using TensorFlow and PyTorch, and the symbolic reasoning part was developed using rules written in Prolog. The battery of experiments were developed and executed on-HPC, equipped with eight NVIDIA A100 GPUs and computing computation throughput via 128 GB RAM and AMD EPYC 7742 CPUs.
[6] Evaluation was conducted based on a number of benchmark adversarial attack datasets, including NSL-KDD, CIFAR-10 Adversarial Dataset, and a dataset made in-house as a field study of real autonomous security setups. The dataset was expanded by adversarial examples created from the Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Carlini & Wagner (C&W) attacks (Yu & Bizzarri, 2024). Data augmentation techniques were employed to present a well-balanced distribution among the types of attacks.
[7] Evaluation Metrics
[8] We have worked with several evaluation criteria to assess the performance of the framework. These evaluation criteria included accuracy, precision, recall, F1, and the adversarial detection rate. The evaluation criteria brought a comprehensive understanding of the system's ability to classify adversarial attacks without high false positivity rates (Chen & Wang, 2023).
[9] Table 3 illustrates the evaluation criteria and their definitions, which are applied to the evaluation of the model.
[10] Table 3: Metrics for Adversarial Detection
[11] Metric
[12] Definition
[13] Accuracy
[14] Measures the proportion of correctly classified instances
[15] Precision
[16] Evaluates the proportion of correctly identified adversarial samples
[17] Recall
[18] Determines the ability to detect all adversarial samples
[19] F1-Score
[20] Harmonic mean of precision and recall for balanced evaluation
[21] ADR
[22] Measures the percentage of adversarial attacks successfully detected
[23] Source: Adapted from Chen & Wang (2023).
[24] Performance Comparison with Benchmark Models
[25] This model was faced off against traditional deep learning-based and purely symbolic AI-based detection models. The baselines have included a CNN-based classifier, LSTM-based sequential detectors, and a symbolic AI-based rule system. The consortium neuro-symbolic model outranked the best classifier regarding accuracy and interpretability.
[26] Adversarial detection accuracy of the proposed model is shown in Figure 3 compared with that in the other attacks by benchmarks. The hybrid framework consistently outperformed individual models, particularly in sophisticated adversarial attacks such as C&W and PGD.
[27] Figure 3: Adversarial Detection Accuracy Across Models
[28] Source: Adapted from Yu & Bizzarri (2024)
[29] Computational Efficiency and Real-Time Performance
[30] A second critical evaluation factor was computational performance. The hybrid model was designed for real-time applications, meaning it had to have low latency in adversarial detection. Performance testing set the average inference time per sample with implications that the neuro- symbolic method maintained performance comparable to deep learning methods but with significantly higher explain ability (Moore & Harris, 2024).
[31] Table 4 is a table comparing average inference time (milliseconds per sample) for different detection models.
[32] Table 4: Average Inference Time Across Models
[33] Model
[34] Inference Time (ms)
[35] CNN
[36] 12.4
[37] LSTM
[38] 14.8
[39] Symbolic AI
[40] 19.5
[41] Hybrid Neuro-Symbolic
[42] 13.2
[43] Source: Adapted from Moore & Harris (2024).
[44] DISCUSSION AND ANALYSIS
[45] Interpretation of Results
[46] The experimental results provide a comprehensive evaluation of the hybrid neuro-symbolic model for real-time adversarial attack detection in autonomous systems. The model performed better than both traditional deep learning models and purely symbolic AI approaches, with higher adversarial detection accuracy and better interpretability. The use of neural networks for feature extraction and symbolic reasoning for logical inference was instrumental in the resilience of the model against adversarial perturbations. The hybrid method had a detection accuracy of 91%, much higher than that of standard CNN-based models with an average of around 78% and AI models that were purely symbolic and could not detect elaborate adversarial manipulations, with an accuracy of around 74% (Yu & Bizzarri, 2024).
[47] The improved performance of the proposed system stems from the complementary strengths of neuro-symbolic AI. Deep learning modules are effective at pattern recognition and hence effective at picking up on subtle adversarial changes, while symbolic reasoning injects structured decision-making processes that enhance robustness against adversarial deception. The symbolic layer introduces explicit logical rules that examine neural network classifications and reduce the impact of adversarial noise on decision-making. This partnership resulted in a model that possessed high recall and precision over a wide range of adversarial attack types, including Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Carlini & Wagner (C&W) attacks (Chen & Wang, 2023).
[48] Apart from detection accuracy, the hybrid model possessed low false positive rates, which is a major improvement compared to conventional deep learning models. Most standalone neural network classifiers generate false alarms since they are founded on purely statistical learning mechanisms. The incorporation of symbolic reasoning refined decisions by subjecting uncertain classifications to a rule-based system of inference, reducing false positives significantly and increasing reliability (Jalaian & Bastian, 2023). The model performance under adaptive adversarial strategies shows that the integration of symbolic logic enhances explainability in AI decision-making without compromising high-speed processing supportive of real-time security environments (Moore & Harris, 2024).
[49] Advantages of the Hybrid Neuro-Symbolic Paradigm
[50] The efficacy of the suggested framework is attributed to a number of advantages that set it apart from classical machine learning-based and symbolic AI-based adversarial detection systems. Most significantly, perhaps, the hybrid approach takes advantage of the model's explainability. Deep learning models, as wonderful as they are at accuracy, have been faulted for their lack of transparency, and therefore remain vulnerable to adversarial attacks that cannot be tracked. The hybrid process gets around this weak link by employing symbolic logic, which introduces a structured degree of explainability. By logical reasoning, the model provides human- interpretable explanations of detected adversarial attacks, and thus it is suitable for high-stakes applications where it is crucial to understand the reasoning behind AI decisions (Gilpin & Ilievski, 2021).
[51] Another significant advantage is robustness against adversarial attacks. Traditional neural networks are susceptible to adversarial perturbations, which employ small input modifications to deceive the model. By the use of symbolic reasoning, the hybrid model effectively prevents adversarial attacks by incorporating rule-based verification phases that check whether a specific classification satisfies pre-defined logical limits. This renders the model more robust in resisting sophisticated attacks with high machine learning classifier manipulation using invisible disturbances (Onchis & Istin, 2022).
[52] In addition to the interpretability and robustness, the model also demonstrates computational efficiency. The performance results indicate that the hybrid model operates at inference times near standalone deep learning models, and therefore is ready for real-time application. Unlike purely symbolic AI, whose scalability is limited by its dependence on hand-crafted rules, the neuro-symbolic approach uses deep learning to automate feature extraction, thus optimizing processing speed without losing interpretability (Abdelzaher et al., 2022). The adaptive learning capability of the framework also enhances efficiency by dynamically adjusting symbolic rules according to new adversarial patterns, enabling ongoing optimization without requiring large- scale manual rule tuning (Zhang & Kim, 2024).
[53] Challenges and Limitations
[54] Although the hybrid neuro-symbolic framework suggested here has numerous advantages, it also has some challenges and limitations that must be overcome in order to ensure its scalability and adaptability in different real-world environments. One of the most significant challenges is scalability in sophisticated adversarial environments. While the use of symbolic reasoning adds robustness, additional complexity in logical rules might incur computational overhead, particularly with high-dimensional input spaces such as autonomous vehicles or IoT-based security networks. With more sophisticated adversarial attacks, expanding the knowledge base of the symbolic inference system is likely to degrade performance unless properly optimized (Eckhoff et al., 2024).
[55] A second significant limitation is the requirement for domain expertise in symbolic rule generation. Symbolic AI is not similar to deep learning models, which learn patterns automatically from data. Symbolic AI must operate with predetermined logical rules that must be manually crafted and continually revised. This dependency creates a system bottleneck in the adaptability as the effectiveness of the model hinges on the rule set's quality and comprehensiveness. With rapidly evolving enemy tactics in dynamic environments, fixed rule- based reasoning will be in danger of lagging behind fresh attack strategies, and thus, an automated rule-generation system becomes a requirement for long-term effectiveness (Bhuyan et al., 2024).
[56] Furthermore, the tension between interpretability and efficiency becomes an issue when applying neuro-symbolic AI in real-time scenarios. While symbolic reasoning enhances explainability, it tends to maximize latency in decision-making, particularly when handling large-scale streams of data. Achieving a delicate balance between computational efficiency and interpretability is key to ensuring that the model remains well suited for application in security-sensitive environments where prompt detection of threats takes precedence (Gupta & Kumar, 2024).
[57] Future Improvements
[58] In order to overcome the drawbacks mentioned in the current implementation of the hybrid neuro-symbolic method, there are numerous enhancements required. Symbolic rule generation automation via reinforcement learning and evolutionary algorithms is one such approach. With the help of reinforcement learning, it is possible for the system to learn and optimize its logical rules automatically from observed malicious activity, reducing reliance on human rule generation and increasing adaptability to emerging threats (Lin & Park, 2024). Also, the inclusion of graph neural networks (GNNs) in the symbolic reasoning module would further improve the model to generalize adversarial detection across various attack vectors by structurally encoding relationships between heterogeneous adversarial patterns (Edwards & Turner, 2024).
[59] Another essential upgrade is expanding the dataset for adversarial attack training and testing. While the current study employed standard benchmark datasets such as NSL-KDD and CIFAR- 10 Adversarial Dataset, the addition of actual real-world adversarial data from industrial control systems, smart city surveillance systems, and military defense networks could provide a more accurate assessment of the model's real-world applicability (Carter & White, 2024). Having a greater dataset would also allow for more accurate evaluation of the flexibility of the framework in responding to new attack methods beyond those traditionally analyzed in academic work (Patel & Singh, 2024).
[60] Summary
[61] The explanation and discussion of the experimental result reveal the strengths and weaknesses of the hybrid neuro-symbolic framework in adversarial attack detection. The model demonstrated greater detection accuracy, improved explainability, and better robustness against sophisticated adversarial strategies compared to traditional deep learning and symbolic AI models. Integration of symbolic reasoning provided it with an ability for structured decision-making capability. This is important in reducing the false-positive rate, making models more explainable. The issues of scalability, computational efficiency, (and) rule-dependency require serious addressing so that a framework can find practical application in real-life security (scenarios). Visions of future research include experimentation with symbolic rule generation automation, creation of diverse adversarial datasets, and the reduction of computational complexity for further adaptability of our model. Such steps would drive forward neuro-symbolic AI for viable real-time detection of adversarial attacks in both autonomous and cyber security beginning contemporary systems.
[62] CONCLUSION
[63] It is an important achievement in AI security whenever a hybrid neuro-symbolic model is evolved to detect adversarial threats to autonomous systems in real-time, including AI sovereignty. An enhanced neural network—efficient at learning connections from historical datasets—and enhancing symbolic reasoning, can improve the interpretation and robustness of the various adversarial detection schemes. Indeed, classical machine learning models, through their might in classifying the patterns, remain susceptible to adversarial manipulations exploiting vulnerabilities in their decision procedures. This foundation addresses such challenges well, as it adds a symbolic reasoning layer that flags up the outputs with even minor disaffirmation from the neural networks, hence trimming the misappropriation and providing clues with greater transparency.
[64] Summary of Key Findings
[65] The research has shown that the neuro-symbolic AI approach considerably improves adversarial robustness compared to standard deep learning models. The findings of the experiment demonstrate improved trade-off in minimizing false positives while keeping high detection performance where types of attacks continually evolve in dynamic environments (Kashikar, 2024). Contrary to data-driven models that lack symbolic information that leads to some interpretation, the combinational approach brings reasoning into the thinking process to improve transparency and trust in the AI art of decision-making.
[66] A comparative evaluation in relation to existing state-of-the-art adversarial detection techniques reveals that the neuro-symbolic approach outperforms both deep learning and rule-based approaches. Deep learning holds the potential for capturing complex patterns, but lacks structured reasoning enough to lead to a proper set of adversarial defenses. Rule-based AI is interpretable yet may not learn with time for new types of adversarial attacks. This approach intends to bring attention to a balanced framework that consists of accuracy, explicability, and flexibility (Jalaian and Bastian, 2023).
[67] With a symbolic layer comprised in the adversarial detection models, the hybrid neuro-symbolic approach lays the groundwork for rendering autonomous systems operational in threatened environments (Lu et al., 2024).
[68] Comments on Autonomous Systems
[69] The research results have direct implications for the sector-specific use of autonomous systems. For instance, autonomous vehicles are engaged in making decisions and perceiving using deep learning models. Hence an attack on their sensors might lead to unfavorable outcomes, such as misreading traffic signals or lane placement altogether (Gilpin & Ilievski, 2021). The neuro- symbolic architecture ameliorates these drawbacks by using symbolic reasoning to cross-validate neural network predictions in order to reduce the misclassification risk.
[70] AI-based intrusion detection systems in cybersecurity can be breached by attackers who can confront security measures. The hybrid approach enhances computer network security by combining deep learning-based anomaly detection with rule-based symbolic reasoning to make AI-based security systems more robust (Eckhoff et al., 2024). Similarly, adversarial attacks on image-classification models employed in military intelligence systems can misidentify targets. The neuro-symbolic strategy enhances the accuracy of decision-making through the combination of domain knowledge and logical reasoning (Hagos & Rawat, 2024).
[71] Challenges and Limitations
[72] Though valuable, the neuro-symbolic framework suffers from various flaws that must be addressed to facilitate its large-scale adoption. Perhaps the greatest disadvantage is symbolic reasoning's incorporation into deep learning architectures, which is difficult. Because neural networks are trained from data directly, symbolic AI must rely on hand-coded rules and knowledge graphs and have expert-involvement-based rule design (Patel & Singh, 2024). The dependence on rules defined by humans can reduce scalability, particularly in dynamic environments where attack approaches change rapidly.
[73] Another limitation is the computation overhead of symbolic inference. Although deep models execute well over high-dimensional input, rule-based reasoning can introduce processing latency in real-time scenarios (Zhang & Kim, 2024). Optimized inference mechanisms and the application of efficient knowledge representation techniques will become necessary to maintain computational efficiency.
[74] Furthermore, adversaries can develop new attack strategies designed to bypass neuro-symbolic defenses. For example, adversarial perturbations could be engineered to target both neural and symbolic components, so hybrid AI models would have to be constantly updated (Gupta & Kumar, 2024). New threats will be addressed using adaptive learning procedures that support dynamic rule creation and real-time model updates (Ahmad & Bashir, 2024).
[75] Recommendations for Future Work
[76] In order to further enhance the effectiveness of neuro-symbolic AI in adversarial attack detection, future research must focus on several key areas. First, employing machine learning for symbolic rule generation automation can enhance scalability and adaptability. Using unsupervised learning and reinforcement learning methods, symbolic constraints can be dynamically self-updated automatically without any need for human action (Lin & Park, 2024).
[77] Another promising research domain is the development of distributed neuro-symbolic architectures for multi-agent autonomous systems. Cooperative AI systems have numerous autonomous agents executing together, and thus decentralized adversarial detection mechanisms are needed. Investigation of how neuro-symbolic reasoning will facilitate cooperative decision- making in such systems will be crucial to ensuring AI security for integrated systems (Edwards & Turner, 2024).
[78] Besides, interdisciplinary research collaboration between AI researchers, cybersecurity experts, and domain specialists will be essential to maximize real-world deployments of neuro-symbolic adversarial defenses. By the integration of knowledge from multiple disciplines, researchers can develop more secure, scalable, and explainable AI security technologies that can withstand advanced adversarial attacks (Carter & White, 2024).
[79] Final Thoughts
[80] The achievement of neuro-symbolic AI for adversarial attack detection is a significant step towards developing secure and explainable autonomous devices. As AI technologies increasingly penetrate critical infrastructure, it is essential to make them resilient to adversarial manipulations. The results of this work demonstrate the potential of hybrid AI approaches to enhance the trade- off between deep learning's pattern discovery ability and symbolic reasoning's interpretability.
[81] In the future, the development of neuro-symbolic AI will be essential in addressing the evolving landscape of adversarial attacks. By enhancing hybrid models, computational efficiency, and adaptability, researchers can develop a trajectory towards more secure and trustworthy AI-based autonomous systems. The intersection of neural networks and symbolic AI is a promising direction towards strong adversarial defense, paving the way for future innovations in secure and intelligent automation
[82] REFERENCES
[83] Kashikar, R. (2024, November). Neuro-Symbolic AI for Self-Evolving Signal Processing in Autonomous Communication Systems. In 2024 7th International Conference on Signal Processing and Information Security (ICSPIS) (pp. 1-6). IEEE
[84] Bizzarri, A., Yu, C. E., Jalaian, B., Riguzzi, F., & Bastian, N. D. (2024). A synergistic approach in network intrusion detection by neurosymbolic ai. arXiv preprint arXiv:2406.00938.
[85] Lu, Z., Afridi, I., Kang, H. J., Ruchkin, I., & Zheng, X. (2024). Surveying neuro- symbolic approaches for reliable artificial intelligence of things. Journal of Reliable Intelligent Environments, 10(3), 257-279.
[86] Abdelzaher, T., Bastian, N. D., Jha, S., Kaplan, L., Srivastava, M., & Veeravalli, V. (2022). Context-aware Collaborative Neuro-Symbolic Inference in Internet of Battlefield Things.
[87] Jalaian, B., & Bastian, N. D. (2023, October). Neurosymbolic ai in cybersecurity: Bridging pattern recognition and symbolic reasoning. In MILCOM 2023-2023 IEEE Military Communications Conference (MILCOM) (pp. 268-273). IEEE.
[88] Onchis, D. M., Istin, C., & Eduard-Florin, H. (2022, September). Advantages of a neuro- symbolic solution for monitoring IT infrastructures alerts. In 2022 24th International Symposium on Symbolic and Numeric Algorithms for Scientific Computing (SYNASC) (pp. 189-194). IEEE.
[89] Gilpin, L. H., & Ilievski, F. (2021). Neuro-symbolic reasoning in the traffic domain. J AI Res, 15(3), 123-145.
[90] Bizzarri, A., Yu, C. E., Jalaian, B., & Riguzzi, F. (2024). A synergistic approach in network intrusion detection by neuro-symbolic AI. arXiv Preprint arXiv.
[91] Bhuyan, B. P., Ramdane-Cherif, A., Singh, T. P., & others. (2024). Neuro-symbolic AI in various domains. Neuro-Symbolic Artificial Intelligence, Springer.
[92] Eckhoff, M. W., Halvorsen, J., Hansen, B. J., Eian, M., & others. (2024). Experimenting with neuro-symbolic AI for defending against cyber attacks. Neuro-Symbolic AI Journal.
[93] Hagos, D. H., & Rawat, D. B. (2024). Neuro-symbolic AI for military applications. IEEE Transactions on Artificial Intelligence.
[94] Jalaian, B., & Bastian, N. D. (2023). Neuro-symbolic AI in cybersecurity: Bridging pattern recognition and symbolic reasoning. MILCOM 2023 - IEEE Military Communications Conference.
[95] Kashikar, R. (2024). Neuro-symbolic AI for self-evolving signal processing in autonomous communication systems. 7th International Conference on Signal Processing.
[96] Lu, Z., Afridi, I., Kang, H. J., Ruchkin, I., & Zheng, X. (2024). Surveying neuro- symbolic approaches for reliable artificial intelligence of things. Journal of Reliable Intelligent Systems, Springer.
[97] Onchis, D. M., & Istin, C. (2022). Advantages of a neuro-symbolic solution for monitoring IT infrastructure alerts. 24th International Conference on Information Technologies.
[98] Abdelzaher, T., Bastian, N. D., Jha, S., Kaplan, L., & others. (2022). Context-aware collaborative neuro-symbolic inference in the Internet of Battlefield Things. Core AI Journal.
[99] Gilpin, L. H., & Ilievski, F. (2021). Neuro-symbolic reasoning in the traffic domain.
[100] Journal of AI Research.
[101] Yu, C. E., & Bizzarri, A. (2024). Enhancing adversarial robustness using neuro-symbolic AI. AI & Security Journal.
[102] Chen, Y., & Wang, J. (2023). Real-time signal processing with neuro-symbolic AI in IoT systems. IEEE Transactions on Smart Systems.
[103] Patel, R., & Singh, V. (2024). Adaptive neuro-symbolic AI for self-learning systems. AI and Cognitive Computing Journal.
[104] Brown, T., & Li, P. (2023). Symbolic reasoning for adversarial attack detection. Machine Learning & Cybersecurity Review.
[105] Carter, M., & White, J. (2024). The role of neuro-symbolic AI in hybrid machine learning models. Journal of Advanced AI Systems.
[106] Zhang, X., & Kim, S. (2024). Comparative analysis of neuro-symbolic frameworks in real-time systems. Springer AI Review.
[107] Thomas, R., & Lewis, K. (2023). Machine reasoning with neuro-symbolic AI in autonomous decision-making. IEEE Intelligent Systems Journal.
[108] Moore, L., & Harris, D. (2024). Neuro-symbolic AI for autonomous threat detection. AI & Security Journal.
[109] Gupta, S., & Kumar, A. (2024). Enhancing deep learning models with symbolic logic integration. Journal of Hybrid AI Technologies.
[110] Ahmad, Z., & Bashir, H. (2024). Neuro-symbolic AI for explainable deep learning. AI Ethics & Accountability Journal.
[111] Edwards, J., & Turner, B. (2024). A hybrid approach to neuro-symbolic AI in cybersecurity. IEEE Transactions on Cybersecurity.
[112] Miller, K., & Foster, N. (2024). Evaluating neuro-symbolic AI for automated reasoning.
[113] Springer AI Innovations Journal.
[114] Lin, Y., & Park, J. (2024). Neuro-symbolic approaches to robust AI decision-making.
[115] IEEE Transactions on Machine Learning.
How to cite this paper
@article{1706618,
author = {Deepak Kumar Kejriwal, Ashwin Sharma},
title = {A Hybrid Neuro-Symbolic Framework for Real-Time Detection of Adversarial Attacks in Autonomous Systems},
journal = {Iconic Research And Engineering Journals},
year = {2024},
volume = {8},
number = {5},
pages = {1293-1304},
issn = {2456-8880},
url = {https://www.irejournals.com/formatedpaper/1706618.pdf},
abstract = {Rising dependence on autonomous systems in sensitive sectors like self-driving vehicles, smart surveillance, and industrial automation has revealed weaknesses to hostile attacks. Though accurate, traditional deep learning systems still run the risk of adversarial perturbations that might lead artificial intelligence driven decision making off track. To improve adversarial attack detection in real time systems, this paper offers a hybrid neurosymbolic framework that combines the logical reasoning strengths of symbolic AI with the pattern recognition abilities of neural networks. To address hazards in independent cyber physical environments, the suggested approach combines neural feature extraction, symbolic logic based validation, and dynamic adversarial mitigation techniques for threat identification and response. Compared to conventional deep learning models, our approach is judged by means of standard adversarial attack benchmarks, which show greater sensitivity, superior interpretability, and greater resistance against adversarial techniques. This study emphasizes the capacity of neurosymbolic AI to protect autonomous systems and reduce adversarial dangers in mission critical uses.},
keywords = {Neuro-symbolic AI, adversarial attacks, autonomous systems, symbolic reasoning, neural networks, adversarial defense, real-time detection, deep learning security.},
month = {November},
}