This post is also available in:
עברית (Hebrew)
Written by Or Shalom
Effective defense against threats such small, low-speed unmanned aerial systems (LSS – Low, Slow, Small), requires a multidimensional approach. Modern detection architectures are divided into the following technological domains: the electromagnetic domain (radar), the electro-optical and thermal domain (EO/IR), the acoustic domain, and the electronic domain (RF/EW). The logic behind multi-layered defense lies in the principle of Failure Compensation. Since LSS threats are specifically designed to exploit the weaknesses of each standalone sensor (for example, low radar cross-section or GPS disruption), only parallel, complementary data fusion across multiple sensors can ensure reliability, resilience, and comprehensive operational coverage. AI implementation is the key to transforming this complex set of sensors and the abilities into one coherent, fast-acting system.
Standard detection systems struggle with sophisticated threats, such as LSS, due to three core challenges: first is the physical dimension (evasion), in which small, slow targets exploit low RCS, minimal thermal signatures, and low-altitude flight profiles to evade single-sensor detection; second is the data dimension (noise), in which heterogeneous sensors generate conflicting or partial information, and significant environmental noise (such as bird flocks or civilian drones), which creates high false-positive alarm rates that undermine confidence in the system; third is the operational dimension (response time), in which the need for rapid detection, classification, and response demands decision-making speeds beyond human capability [1].
AI as the Enabler of Multi-Sensor Fusion:
AI’s central value lies in its ability to perform reliable data fusion across heterogeneous sources, transforming a layered system into a single operational unit. This manifests in two key mechanisms:
- Advanced data fusion (Low-Level Fusion): Instead of comparing only the final outputs of sensors, AI can fuse raw signal streams before generating a Track. This enables the detection of extremely weak signatures that no single sensor could identify alone, but which collectively form a reliable threat vector.
- Noise filtering and false-alarm reduction: Machine-learning algorithms learn to differentiate legitimate UAS activity from natural or electronic noise. This dramatically reduces operator load and strengthens trust in the system.
AI Capabilities for Managing Uncertainty:
AI’s most critical contribution to detection systems is its ability to operate effectively under partial or conflicting information using several key mechanisms:
- Confidence and Trust: Uncertainty Quantification (UQ) – A successful AI system must know not only what it thinks, but also how confident it is in its prediction. This is the principle of Uncertainty Quantification (UQ). This capability is critical since it prevents models from developing systemic overconfidence and delivering high predictions for corrupted or unfamiliar data. The AI measures two main dimensions of uncertainty (the model’s lack of knowledge, and noise in the data). Based on these measurements, it ensures that critical actions are only executed when both the data and the prediction are reliable. The core operational advantage is that in situations of high uncertainty, the system automatically triggers an Escalation Procedure and hands over the decision to a human operator, thereby enabling the Risk-Adjusted Decision-Making and maintaining human control precisely at the critical moments [2].
- One of the most promising core capabilities of advanced AI is Out-of-Distribution (OOD) Detection. A well-trained AI system can autonomously recognize when it encounters an object or a signal that had never been seen in its original training dataset (such as a new drone model, an unfamiliar jamming technology, or an anomalous electronic signature). Instead of making an erroneous guess, the model utilizes high uncertainty (as described above, in Uncertainty Quantification) as the primary metric indicating that the threat is unfamiliar. The decisive operational advantage is that the system provides an early, data-driven alert about the emergence of new adversary technology. This enables EW (Electronic Warfare) or intelligence systems to swiftly allocate resources for the rapid and accurate analysis of the unknown threat, thereby dramatically shortening the decision loop against a modern and adaptive opponent.
- The third and critical capability of advanced AI is Continuous Learning and Dynamic Adaptation. This insight means the AI system rapidly learns from new threats; after every detection event of a novel threat (as flagged by OOD), the system automatically updates its internal models by rapidly incorporating the characteristics of the newly gathered threat data. This mechanism ensures that the operational knowledge is always current. The distinct operational advantage is maintaining constant operational relevance and dramatically reducing the knowledge gap against a modular and adaptive adversary. This automatic adaptation capability ensures that the system improves its performance and accelerates its response time exponentially in every future encounter with the previously recognized threat [3].
Integrated Example:
In a scenario where a detection system encounters an unmanned aircraft (UAV) with a never-before-seen signature, all three AI principles operate seamlessly:
- OOD Detection: When the model appears, the system compares the radar’s data and electronic signature to the existing database. Instead of classifying it as a mistake, the system identifies a non-matching signature and flags it as an OOD, and activates the “new and unknown threat” alert.
- UQ (Uncertainty Quantification)-based escalation: Since the threat was flagged as OOD, the system automatically calculates that the uncertainty is very high. Instead of acting autonomously, the AI activates an Escalation Procedure and relays the raw data, along with the “highly uncertain” label, to the operations room. There, a senior human operator carries out the Risk-Adjusted Decision-Making.
- Continuous learning: After successfully handling the threat, all the information regarding the model is collected, and the AI system incorporates the new information in its database and automatically updates the models, ensuring future encounters are detected with high confidence (low UQ), and enabling a continuous autonomous response to any tactical innovations introduced by the enemy.
Through these mechanisms, AI enables multi-layered defense systems to maintain operational relevance and resilience against evolving, learning adversaries.
The author is a security, cyber and HLS technology expert and consultant to government ministries and defense industries. He holds a master’s degree, as well as civil and national qualifications in the realm of HLS and Cyber Security. He has experience in consultation and business development for security companies and groups in matters of planning and building defense, innovation and security technology, exercises, and training in security and cyber.
[1] https://www.mdpi.com/2072-4292/16/5/879
