The review traces the IoT’s evolution through three distinct generations. The first consisted of passive sensors and actuators that simply fed data to centralized control units, with every decision made by humans. The second, often called Smart IoT, brought context awareness by coupling devices with cloud platforms, edge computing, and more capable embedded processors, enabling limited local decisions in smart homes and connected vehicles. The third generation, A-IoT, marks a genuine paradigm shift: devices powered by artificial intelligence, distributed edge computing, and high-speed 5G and 6G links that learn from data, make predictions, and execute real-time actions on their own. In precision agriculture, autonomous drones now analyze crop health and initiate spraying without human intervention; self-driving cars negotiate with roadside infrastructure to adapt to shifting traffic conditions.

That autonomy, however, comes at a steep security price. A-IoT networks are adaptive, open-ended, and often ad hoc in structure, and the mechanisms that enable autonomy—self-learning models, continuous inter-device communication, and collaborative behavior—simultaneously widen the attack surface. The review catalogs a formidable set of threats: adversarial machine learning that corrupts AI models by manipulating training data or injecting malicious inputs during inference, data poisoning, spoofing, botnet propagation, firmware tampering, and exploitation of zero-day vulnerabilities. Because most A-IoT deployments lack centralized control and a unified security policy, real-time threat detection and coordinated mitigation become extraordinarily difficult. The problem is compounded by the heterogeneity of the ecosystem, where billions of nodes with different hardware architectures, operating systems, protocols, and roles must be authenticated and protected without central management.

Resource constraints add another layer of difficulty. Most edge devices—sensors, actuators, embedded controllers—operate with limited computation, memory, and energy, which rules out heavyweight encryption and computationally intensive machine learning models. At the same time, A-IoT is fundamentally real-time: decisions must be made within milliseconds to keep autonomous driving, remote surgery, and industrial automation safe. Intrusion detection and anomaly detection systems therefore need to be both fast and precise, with low false-positive rates, so they do not disrupt time-sensitive operations. The authors also highlight the risk of data leakage: autonomous devices continuously collect sensitive information, from personal health data to geolocation, and their decentralized, mobile nature increases the chances of interception, alteration, or exfiltration during transmission or storage—challenges that traditional Public Key Infrastructure and blockchain solutions may struggle to meet on constrained edge hardware.

The core of the review maps the AI toolkit onto these defensive tasks. Supervised machine learning excels at malware classification and intrusion detection where labeled datasets of known threats exist, while unsupervised techniques such as k-means clustering, autoencoders, and one-class support vector machines flag anomalous behavior without labels—an essential capability for spotting novel or evasive attacks in environments where normal behavior is itself situational. A drone that unexpectedly deviates from its flight path or alters its communication protocol, for example, may be signaling a cyber-physical attack. Deep learning, including convolutional and recurrent neural networks, extracts hierarchical features from raw network traffic and can detect encrypted malware and profile node behavior with high accuracy, though it demands substantial data and compute power that resource-limited devices cannot always supply.

Federated learning emerges as a privacy-preserving alternative that suits the decentralized character of A-IoT. Rather than shipping raw data to a central server, devices collaboratively train a global model and share only local model updates, protecting sensitive data while reducing communication overhead—a design particularly attractive for smart healthcare and industrial systems. Swarm intelligence, inspired by the collective behavior of ants and birds, offers another distributed approach: in a swarm-based intrusion detection system, devices exchange anomaly scores and coordinate responses in a decentralized fashion, making the overall defense robust to local failures or targeted assaults. Such tactics are especially valuable in mobile or mission-critical contexts like drone swarms, battlefield networks, and disaster recovery operations.

The review also spotlights emerging paradigms that could reshape the field. Explainable AI addresses the black-box problem, revealing which parameters—packet frequency, device location, protocol behavior—influenced an intrusion detection decision, thereby building trust and aiding compliance with regulations such as GDPR and NIST standards. Generative adversarial networks can simulate realistic zero-day and polymorphic attack patterns, allowing defensive systems to train in richer threat environments and enabling automated red-teaming. Transfer and meta-learning let pre-trained models be rapidly fine-tuned on small, domain-specific datasets, reducing training time at the edge. Digital twins—real-time virtual counterparts of physical systems—allow defenders to simulate hypothetical cyberattacks on autonomous vehicles or smart manufacturing plants, supporting scenario-based training, resilience testing, and proactive risk management without touching live infrastructure.

Real-world deployments are already validating these ideas. Smart cities are testing federated learning-based intrusion detection that processes data at edge nodes to keep sensitive information local. In automotive security, anomaly detection algorithms monitor Controller Area Network bus operations to catch unauthorized command injections, while deep reinforcement learning optimizes drone flight routes in threat scenarios. Industrial operators deploy AI-enhanced Security Information and Event Management systems that learn anomalous patterns in industrial control and SCADA environments, and generative AI is used to model advanced persistent threats before they strike. Hospitals are adopting federated learning to train diagnostic models across institutions without sharing patient data, and energy utilities apply AI-based anomaly detection to millions of smart meters, sometimes reinforced with blockchain for tamper-resistant data provenance.

Yet the review is candid about the gaps. AI models themselves are vulnerable: adversarial inputs can silently mislead classifiers, and poisoning attacks contaminate training data. Explainability techniques remain computationally expensive and immature for real-time security applications. Labeled attack data—especially for zero-day threats—is scarce, fueling interest in GAN-generated synthetic datasets. There is also a fundamental trade-off between security and performance: stronger protection generally means more computation, memory, and communication overhead, straining battery life and latency on lightweight devices. Beyond the technical realm, fragmented legal frameworks, unresolved questions of algorithmic bias and accountability, and the absence of universal standards for AI model validation and interoperability impede coherent, cross-vendor defense strategies.

Looking ahead, the authors chart several promising research directions. Hybrid AI-blockchain frameworks could combine immutable, decentralized ledgers for secure data sharing with AI-driven anomaly detection and trust management, provided lightweight consensus mechanisms can fit within IoT resource budgets. Self-healing security agents—systems that continuously monitor network health and automatically recover from breaches without human intervention—represent an ambitious frontier, demanding careful balance between autonomy and control. AI-powered threat intelligence platforms, multi-modal detection that fuses network signals, sensor readings, and user behavior analytics, and eventually fully autonomous cybersecurity agents operating across scattered networks complete the vision. The review’s overarching message is clear: securing the autonomous digital frontier will require adaptive, transparent, and ethically governed AI, developed through interdisciplinary collaboration across AI, cryptography, embedded systems, and real-time engineering, so that the networks of the future can defend themselves as intelligently as they operate.

Subject of Research: AI-driven cybersecurity strategies for autonomous Internet of Things ecosystems

Article Title: Securing the future: AI-driven cybersecurity in the age of autonomous IoT

Article References: Ogenyi, F. C., Ugwu, C. N., & Ugwu, O. P.-C. (2025). Securing the future: AI-driven cybersecurity in the age of autonomous IoT. Frontiers in the Internet of Things, 4, Article 1658273. https://doi.org/10.3389/friot.2025.1658273

Image Credits: AI Generated

DOI: 10.3389/friot.2025.1658273

Keywords: autonomous IoT, AI-driven cybersecurity, intrusion detection, federated learning, explainable AI, generative adversarial networks, digital twins, adversarial machine learning, zero-day vulnerabilities, swarm intelligence, edge computing, blockchain