Dual-Layer Intelligence: Blending Basic Mathematics and Artificial Intelligence for Enriched Threat Detection Systems

Threat Detection in the Defence and Security Industry

The Defence and Security Industry covers a wide range of problem areas. Several companies in the industry focus on detecting threats among aircrafts. Some focus on threats among naval ships. Others focus on threats among land vehicles. Given the diverse range of issues that this industry encounters, it is easy for even the most experienced professionals to become overwhelmed with choosing the right algorithms and models to tackle the problems that may arise.

This article guides readers on when to choose basic Mathematics vs Artificial Intelligence for Defence and Security Industry projects, even using ensemble approaches for a polished threat detection system.

Yes, Artificial Intelligence is Mathematical, but with a Caveat

To be very clear, this article does not make the claim that Artificial Intelligence is isolated from Mathematics. Artificial Intelligence is an insanely broad field that consists of Machine Learning and Deep Learning which apply mathematical concepts such as Linear Algebra, Statistics and Calculus. The difference is that modern Artificial Intelligence is much more complex than basic mathematical techniques and requires more computational power, time and memory to arrive at optimised conclusions. As such, this article uses the term basic Mathematics to differentiate simpler algebraic, geometric, linear and deterministic calculations from those used in Artificial Intelligence which are more statistical, probabilistic, optimisation-focused and data-driven.

The Over-Engineering Trap

There is a famous saying, “Just because you can, does not mean you should.” With Artificial Intelligence being a hot topic, it is tempting to apply it to any scenario. This practice falls under the wider concept of “over-engineering” where a proposed solution is arrived at with more complex means than necessary. While this practice is good for experimental purposes and testing the capabilities of Artificial Intelligence, in practical scenarios, an experienced engineer must carefully select when to use basic Mathematics vs more advanced Machine Learning and Deep Learning models for a project. Let us explore some real-life scenarios.

The Argument for Basic Mathematics in Threat Detection Systems

Perhaps you were the type of student who dreaded Mathematics. You may have thought, “Linear algebra? Geometry? Calculating the slope of a line? Understanding cartesian space? When will I use this in the real world?” I am pleased to inform you that these concepts are quite useful in the Defence and Security Industry! For example, an aircraft threat detection system may want to determine if tracks follow established airways in order to raise alarms for suspiciously wandering aircrafts. This track to airway mapping result may be calculated as an estimated probability of the track following the closest airway. The algorithm may convert the geodetic track points and established airway trajectory points to cartesian space, compare the angles between the slopes of the track trajectory and airway trajectories, compare the distance between the track and airways and calculate any divergence that the track makes from the airways in order to arrive at a final weighted score. Why choose basic Mathematics over Artificial Intelligence in this scenario? Track to commercial airway mappings are not reliant on “cases”, or in technical terms, training data. Artificial Intelligence saves models that are trained on data for specific cases. For example, one may train a model to classify a track as anomalous based on regular and irregular patterns of what is expected of a standard airliner. Airway mappings are not based on a variety of “cases” but on fixed established airway trajectory points saved in a system. As such, relying on computational heavy Artificial Intelligence would be overkill in this scenario which may be achieved with simpler Mathematics that produces deterministic but accurate results.

The Argument for Artificial Intelligence in Threat Detection Systems

Conversely, when the system must interpret irregular or unpredictable behaviour, Deep Learning becomes far more suitable. We will follow up on the previous example of classifying a track as anomalous based on irregular behaviour. This scenario is case dependent as our system needs a rich data set to establish what is “normal” track behaviour and “abnormal” track behaviour. A standard is set; say the system’s standard for normal behaviour reflects that of commercial airliners. We may provide a rich dataset of tracks that exhibit regular values for turn rate, speed and vertical speed according to the standards of regular airliners, labelling these tracks as normal.

Conversely, we may gather tracks that exhibit values that deviate from the standard of the aforementioned characteristics, labelling these tracks as abnormal. A model may be trained with this data to detect abnormalities in incoming flights. This scenario provides an interesting use case for Artificial Intelligence. One may establish this module as a time-series classification problem. The engineer may get creative in experimenting with different neural networks such as ensemble Convolutional Neural Networks (CNNs) as implemented in InceptionTime or Long Short-Term Memory (LSTM). The engineer may train multiple models, fine-tune the hyperparameters for these models, establish loss functions and compare metrics such as accuracy, precision, recall and F1-score in order to choose the most optimal model.

Another scenario to apply Artificial Intelligence is in image-heavy modules. Computer Vision is notorious for its complex nature and heavy computational and memory requirements. It is safe to say that approaching this area with basic Mathematics would prove to be difficult. Let us provide the scenario where a system aims to analyse suspicious behaviour of war tanks captured on video. An optimally trained deep learning model, say a highly precise object detection model, may be used to analyse and classify frame-by-frame tank behaviour irregularities or the presence of unexpected objects captured on film. Although the computation is heavy, this is a scenario where Artificial Intelligence would be ideal.

You Too Can Use Both Approaches

While Artificial Intelligence is certainly a powerful tool for complex tasks, it will not - at least for now - entirely replace the need for simpler mathematical models. A good engineering team combines both approaches and selects the most appropriate scenarios for their applications. This dual perspective allows engineering teams to reserve their resources, namely computational time, power, and memory, for their heaviest, most complex and most demanding modules without compromising on accurate threat detection estimations. Hopefully this article inspires you to use dual basic mathematical and contemporary intelligent systems in your next Defence and Security project!