The Use of Artificial Intelligence in Military Cybersecurity
- 6 hours ago
- 13 min read
Authors: SaiUjwala Mattupalli
Mentor: Antoine Thibault. Antoine is currently a doctoral candidate in the High-Speed Turbomachinery Group in the Department of Engineering at the University of Oxford.
Abstract
In this review paper, the use of artificial intelligence in cybersecurity attack and defense methods is discussed. Military data contains extremely sensitive information which is vulnerable to cybersecurity attacks. It is therefore crucial to study the methods used in cybersecurity attacks to guarantee public safety. The attack methods reviewed include modifying collected data, model inversion attacks and the use of artificial intelligence to spread misinformation. Defense methods are also reviewed. The defense methods considered are collecting large data sets to identify patterns in cybersecurity attacks and cybersecurity decision making tools. This review paper finds that a variety of cybersecurity methods employing artificial intelligence have been used in military applications. The different methods are discussed in case studies in Pakistan and Russia.
Introduction
Artificial intelligence is often used in cybersecurity attacks and defense methods due to its versatility and its ability to make autonomous decisions. Cybersecurity attack methods include techniques to extract sensitive data from the output of a trained artificial intelligence model (model inversion attacks, membership inference attacks), techniques to alter input data used to train artificial intelligence models (poisoning) and data extraction. Cybersecurity attacks can also be used to spread misinformation. To counter cybersecurity attacks, cybersecurity defense methods are also employing artificial intelligence. Artificial intelligence is used in cybersecurity defense methods to collect data and analyze patterns in harmful software, take autonomous security decisions and find patterns in network traffic. The frequency of different cybersecurity attacks was assessed by Rahman et al. (2023). The attack frequencies are presented in Figure 1. The figure shows the percentage frequencies of adversarial, model inversion, poisoning, data poisoning, data extraction, differential privacy and membership inference attacks. Differential privacy attacks refer to adding random noise to data (Rahman et al., 2023). The other attack methods are discussed in greater detail in the literature review.
Figure 1: Percentage frequency of cybersecurity attacks (Rahman et al., 2023).
Adversarial, model inversion and poisoning attacks were identified as the most frequent cybersecurity attacks. Cybersecurity is a crucial aspect of military operations. This involves the protection of sensitive information, espionage and the spread of information at a large scale. Artificial intelligence is increasingly used in these aspects. In this study, the use of artificial intelligence in cybersecurity attack, and defense is reviewed in detail with case studies.
Method
The articles reviewed in this paper include review and research articles. The articles were peer reviewed, ensuring quality in the research covered. The articles reviewed included details in the cyberattack and defense methods, as well as case studies illustrating the use of cybersecurity employing artificial intelligence in a military context. Google Scholar was used as a search engine to find the majority of the papers discussed in the literature review. The keywords used in the literature review search include military cybersecurity, artificial intelligence and malware.
Literature Review
Literature is now reviewed on the attack and defensive cybersecurity methods which employ artificial intelligence. The purpose of this literature review is to discuss the capacity of the novel yet swiftly advancing technology and spread awareness of its strengths and weaknesses.
Cybersecurity attack methods using artificial intelligence
Artificial intelligence features in cybersecurity attacks both as the target of the attack and as the instrument of it. In this section, cybersecurity attack methods involving artificial intelligence are reviewed in both of these forms.
Adversarial attacks
Adversarial attacks include the manipulation of inputs to mislead systems. Such attacks can involve disturbing training samples that in turn causes the system to produce inaccurate responses.
Rahman et al. (2023) suggested that adversarial attacks are the most widely used attacks on machine learning and artificial intelligence models. Such attacks can be carried out in a targeted or untargeted manner. In a targeted attack, the system will result in a specified incorrect output, or a generally incorrect output when untargeted.
Das et al. (2025) further described two distinct subcategories of adversarial attacks; the backdoor attack and the data poisoning attack. Such attacks manipulate training samples to cause the model to produce inaccurate responses. This can diminish system safety, reduce trust placed on the model, and cause real world harm.
Model inversion attacks
Model inversion attacks can be performed on artificial intelligence and machine learning models. Such attacks intend to gain enough information from a system output to reconstruct training data.
Rahman et al. (2023) described model inversion attacks to extract training data information from artificially intelligent models. Sensitive model information can be leaked through this attack. Such attacks can potentially affect military systems by reconstructing security systems.
Membership inference attacks
Membership inference attacks aim to identify if a specific data point was used in the training dataset of the model from its output.
Rahman et al. (2023) noted that in membership inference attacks, the attackers can analyze a model's output to understand a specific data point in the dataset used to train the model. Through this attack, sensitive military information can be leaked. Data sanitization, which involves filtering datasets to remove malicious data points, model pruning, which is removing excessive features to reduce machine complexity and ensure smooth operation, and adversarial training, which involves training models with such attacks are all defensive measures.
Poisoning attacks
Military software, devices and services can be subjected to poisoning attacks. A poisoning attack refers to feeding misleading information into a system to reduce its capabilities. A data poisoning attack specifically targets data to lower the performance of cybersecurity tools employing artificial intelligence.
Hu et al. (2021) categorized poisoning attacks into two groups: availability attacks and integrity attacks. For availability attacks, the poisoning data generated by a supporting system and the points where to poison the system is determined by gradient information. Availability attacks will worsen model performance. Integrity attacks only affect specific inputs chosen by the attacker. Poisoning attacks are detrimental to military organizations due to the various weapons that can be accessed to cause large scale destruction. Defensive measures for such attacks include certified defenses, data sanitization and robustness training.
Malatji et al. (2024) later categorized poisoning attacks differently. Poisoning attacks were characterized as offensive or adversarial depending on the intentions of an attacker. The attack can be regarded as offensive artificial intelligence when aimed to gain access into a system and manipulate data, while it can be regarded as adversarial artificial intelligence when the aim is to exploit or deceive the model. Both forms can be detrimental to military systems. In contrast to Hu et al. (2021) who characterized poisoning attacks in terms of the effect an attack has on a model, Malatji et al. characterized poisoning attacks in terms of the intentions of the attacker.
Nobles (2024) discussed data poisoning in the context of security threats that arise from the vulnerabilities of artificially intelligent systems. Modern cybersecurity attacks can target artificially intelligent models with the intent to control or manipulate. Such attacks gain access to the system through data poisoning.
Data extraction attacks
Military artificial intelligence systems can be subjected to data extraction attacks, where outputs from an artificially intelligent model are used to reconstruct stored data.
Rehman et al. (2023) suggested that data extraction attacks are done by attackers with no significant knowledge of the model. The data used to train the model is usually the target of the attacks. The training data can be inferred based on the output the model gives. Sensitive information can be leaked from data extraction attacks, from a military perspective this can include confidential political information, weapon system codes and more, all harmful if exposed.
The use of artificial intelligence to spread misinformation
Artificial intelligence has a generative capability which can be used to spread misinformation at a large scale.
Zia (2021) focused on the disruptive nature of artificial intelligence in hybrid warfare, noting that artificial intelligence can be used in hybrid warfare to carry out terrorist attacks, suicide bombing and more against the military. Artificial intelligence has also been maliciously used to spread fake audio files and misinformation to create political tension, which can lead to military conflict and war.
Anwer (2024) furthered research into the use of artificial intelligence in a political context and discussed the capability of fake news to create political tensions, demonstrating a requirement for moderating news generated from artificial intelligence. The methods to spread misinformation discussed were deepfakes, social media manipulation and cyber espionage. The spread of misinformation was identified to influence public opinions, election outcomes and government stability, including the US 2016 election, the Brexit referendum and the humanitarian crisis in Myanmar. Cybersecurity defense systems enhanced with artificial intelligence were recommended to help identify patterns in misinformation.
Cybersecurity defense methods using artificial intelligence
Cybersecurity defense methods are increasingly using artificial intelligence to counter more sophisticated threats. In this section, the cybersecurity defense methods employing artificial intelligence are reviewed.
Large data collection
Artificial intelligence algorithms rely on large amounts of data to identify malicious patterns and behaviors.
Hu et al. (2021) discussed the different methods used to attack artificial intelligence systems while the systems are collecting data. Data collection was categorized into software and hardware methods. A given example of a hardware data collection attack was sensor manipulation, where a measurement device would be manipulated to record incorrect measurements. Software data collection attacks discussed included faking or biassing the data collected. Methods to defend against data-collecting attacks were also reviewed. For hardware data collection attacks, this includes filtering sensor data. Software data collection attack defenses mentioned include encrypting data and detecting attacks.
The data used to train defensive cybersecurity systems employing artificial intelligence must not be subjected to cybersecurity attacks. Hu et al. (2021) discussed hardware and software data collection attacks and later Maathuis et al. (2025) discussed the datasets used to train military cybersecurity systems using artificial intelligence. Zeek-based cybersecurity datasets were discussed for training systems using artificial intelligence for threat detection. The datasets enabled cybersecurity systems to recognize user behavior and network traffic patterns leading to attacks. The Simulated Control System Communications Dataset was also discussed for training cybersecurity systems to identify unusual military commands and control.
Autonomous reasoning
Autonomous reasoning involves systems using artificial intelligence to assess situations and make decisions independently without human intervention.
Shoaib (2016) discussed the effectiveness of artificially intelligent cyber weapons for identifying and responding to certain threats. Shoaib notes that offensive artificially intelligent cyber technology is steadily advancing and the need for counteractive defensive technology is more urgent than ever. To counter offensive attacks using artificial intelligence, the same artificial intelligence must be used in defensive systems, including military systems that are responsible for large burdens such as protecting entire nations.
Maik and Afridi (2024) discussed the integration of artificial intelligence in military strategies and defense systems. Artificial intelligence was shown to have been adopted by many military organizations across the world due to its capability to respond to threats in real time and carry out autonomous decisions. Cybersecurity systems employing artificial intelligence were discussed as being able to respond faster to threats than traditional systems. Maik and Afridi (2024) demonstrate the capability of artificially intelligent cyber systems to identify and respond to threats effectively.
Erendor (2025) also highlighted the capabilities of artificial intelligence in military technology to make decisions. It was suggested that autonomous systems which employ artificial intelligence are adept at probabilistic reasoning when faced with a problem. They can do so based on what is inputted into the system, through which they can draft an approach to the situation and instantly decide with no human involvement necessary. Such technology is crucial in military systems because in most cases, it is necessary to obtain information and react faster than the enemy. While Shoaib (2016), Maik and Afridi (2024) and Erendor (2025) all demonstrate the capabilities of artificial intelligent systems to make decisions, there is little evidence in the literature that suggests that the relatively quick autonomous decisions made by artificially intelligent systems result in safer outcomes. This suggests future research into comparing decisions made from traditional and artificially intelligent cybersecurity systems.
Anomaly patterns
Military groups can use artificial intelligence to defend against cybersecurity threats by analyzing patterns in networks and malware codes.
Methods to detect unusual network traffic have been a key research area for cybersecurity systems. Michael et al. (2023) discussed the OODA (observe, orient, decide, act) loop as a method to detect unusual network traffic in military campaigns. The method functions by monitoring network traffic (observe). Patterns are then identified (orient). Software can then be developed to detect unusual network patterns (decide). Finally, the software can respond in real time to traffic patterns without human interaction (act).
Ofusori et al. (2024) also studied pattern recognition software but suggested different methods to identify threats. ANN and CNN were identified as two important neural networks used to detect forms of malware and phishing by recognizing patterns from datasets. CNN can detect phishing or malware by identifying spatial patterns CNN can also detect intrusions to the network. One example of CNN usage is image identification for CAPTCHA images. ANN can identify patterns from network traffic data to detect any intrusions. Such networks that can identify malware, phishing and intrusions are crucial for military systems. The detection methods suggested by Ofusori et al. (2024) use neural networks to identify patterns, while Michael et al. (2023) discussed a continuous approach to monitor, identify, decide and act on network activity.
Discussion
In this section, case studies are discussed where artificial intelligence has been used in military cybersecurity attacks. Recommendations are then made for future research.
The Russian military has been increasing its usage of artificial intelligence in cybersecurity attacks (Thornton & Miron, 2020). The Russian military has been utilizing artificially intelligent systems to improve missiles, electronic warfare, command and defense systems. They have also been extensively testing autonomous weapons recently, signifying their strong interest in incorporating artificial intelligence into military systems. In order to successfully do so and protect any artificial intelligence systems used by the military, further research must be conducted on attack and defense methods regarding artificial intelligence so such systems can be employed to the fullest potential while mitigating any potential risks.
Similarly, Pakistan is another nation that must focus on strengthening its military systems due to its vulnerability to cybersecurity attacks (Bashir et al., 2024). Multiple cyberattacks carried out on Pakistan further support this issue. One of the mentioned attacks was done by India, called “Project Hangover”. Said attack’s aim was to gather information regarding security details in Pakistan. A US spying program named PRISM compromised Pakistan’s security and obtained the information of numerous Pakistan officials. Pakistan’s main cybersecurity vulnerabilities were summarized by Bashir et al.. The summary is described in Table 1.
Table 1: Summary of cybersecurity vulnerabilities and effects in Pakistan (Bashir et al., 2024).
Vulnerability | Effect |
NADRA (responsible for maintaining Pakistan’s biometric and demographic population records) | By hacking the NADRA system, population records and other sensitive information can be leaked and substantially impact the national security of Pakistan. |
National power control center (distributes power and generates electricity) | If the power control center is compromised, the system can be exploited with the employment of other instruments of hybrid warfare to create a synergistic effect. This directly impacts electricity distribution and generation for the entire nation. |
Pakistan’s Sui Gas network (distributes natural gas) | This system heavily depends on cyberspace for regulation, so a cyber-attack on the gas network can cause a gas shortage and energy crisis. |
Communication networks (establishes communication at a global distance) | If the four submarine cables used to connect Pakistan to the global internet cables are compromised, Pakistan will be at the risk of data breaches and communication shutdowns. |
Studies show that cyberattacks driven by artificial intelligence prove to be highly lethal and damaging and recognize such attacks as hybrid warfare. Due to the damage done by cyberattacks on Pakistan, it is clear that nations must reinforce their cyber military systems using artificial intelligence to address any future challenges. It is therefore important for researchers to review the current cybersecurity methods employing artificial intelligence to assist nations anticipate future attacks.
The literature reviewed suggests that cybersecurity systems enhanced with artificial intelligence carry both opportunities and challenges. Cybersecurity systems have the capability to respond autonomously to threats, however there is limited literature to suggest the autonomous decisions result in better outcomes than traditional cybersecurity measures. More research is recommended to assess whether an increase in the use of artificial intelligence in military systems results in better cybersecurity.
Conclusion
In this review paper, the use of artificial intelligence in cybersecurity attacks and defenses in military systems was discussed. The attack methods reviewed include adversarial attacks, model inversion, membership inference, poisoning, data extraction and misinformation spreading. Artificial intelligence has also been employed in cybersecurity defense methods to counter cybersecurity threats. The defense methods reviewed include data collection for training defense systems, autonomous reasoning and identifying patterns in network systems and harmful software.
This review demonstrated the increasing use of artificial intelligence in cybersecurity systems. The use of artificial intelligence in cybersecurity was illustrated with two case studies. In the first case study, several examples of the use of artificial intelligence in Russian military systems were reviewed. The use of artificial intelligence in Russian military systems resulted in a reduction in human input and increase in autonomous weapons. The second case study reviewed how cybersecurity attacks could potentially cause damage to Pakistan’s information, power and communication networks. Pakistan was found to be a victim of two previous cybersecurity attacks to obtain military information. This demonstrates the ability of cybersecurity attacks enhanced with artificial intelligence to extract sensitive military information.
Artificial intelligence is a continuously growing field. Research into the capabilities and limitations of artificial intelligence is important to understand and prevent the damage artificial intelligence may have on military infrastructure.
References
Anwer, A. (2024). Cybersecurity and political warfare: The weaponization of information in the digital age. Interdisciplinary Journal of Humanities, Media, and Political Science, 1(2): 90-97.
Bashir, H., Zarish, W., and Malik, R. (2024). The role of AI in hybrid warfare: A case study of Pakistan’s cybersecurity landscape. Global Strategic & Security Studies Review, 9(1): 86-93.
Das, B, C., Amini, M, H., and Wu, Y. (2025). Security and privacy challenges of large language models: A survey. ACM Computing Surveys, 57(6): 1-39.
Erendor, M, E. (2025). Cyber security in the age of artificial intelligence and autonomous weapons. CRC Press.
Hu, Y., Kuang, W., Qin, Z., Li, K., Zhang, J., Gao, Y., Li, W., and Li, K. (2021). Artificial intelligence security: Threats and countermeasures. ACM computing surveys, 55(1): 1-33.
Maathuis, C., and Cools, K. (2025). Digital sovereignty control framework for military AI-based cyber security. ArXiv.
Maik, H., and Afridi, S. (2024). The role of artificial intelligence in modern warfare and international security. Inkwell publishers.
Malatji, M., and Tolah, A. (2024). Artificial intelligence (AI) cybersecurity dimensions: A comprehensive framework for understanding adversarial and offensive AI. AI and Ethics, 5(2): 833-910.
Michael, K., Abbas, R., and Roussos, G. (2023). AI in cybersecurity: The paradox. IEEE Transactions on Technology and Society, 4(2): 104-109.
Nobles, C. (2024). The weaponization of artificial intelligence in cybersecurity: A systematic review. Procedia Computer Science, 239: 547-555.
Ofusori, L., Bokaba, T., and Mhlongo, S. (2024). Artificial intelligence in cybersecurity: A comprehensive review and future direction. Applied Artificial Intelligence, 38(1): 1-46.
Rahman, M, M., Arshi, A, S., Hasan, M, M., Mishu, S, F., Shahriar, H., and Wu, F. (2023). Security risk and attacks in AI: A survey of security and privacy. IEEE: 1834-1839.
Shoaib, M. (2016). AI-enabled cyber weapons and implications for cybersecurity. Journal of Strategic Affairs: 9-37.
Thornton, R. and Miron, M. (2020). Towards the ‘third revolution in military affairs’: The Russian military’s use of AI-enabled cyber warfare. The RUSI Journal. 165(3): 12-21.
Zia, H. (2021). The evolution of artificial intelligence: Implications for cybersecurity and hybrid warfare. Pakistan Journal of Terrorism Research, 3(1).
Comments