Keywords

Introduction

Artificial intelligence (AI) technology has greatly influenced various aspects of modern society [1] and economy [2]. Recent technological advancements have significantly improved decision-making and support systems, as well as autonomous processes, by utilising different types of data, such as text, sound, visual content, and video footage. To provide accurate outcomes, AI models require collection and preparation of a number of representative data that leads to a costly and timely process in terms of hardware and human resources. Alternatively, frugal AI approaches [3] have emerged to address the data necessity issue, having as a main advantage the use of fewer data for training as learning procedures entail very few samples, with the ultimate challenge being the development of a powerful model, capable of continuously learn with minimal human interactions and no intervention of experts.

Hence, the usage level of an AI model to be exploited under certain conditions is application-oriented as most models comprise supervised approaches and are data-driven. Within the defence domain [4], data may be characterised as limited or incomplete, as well as sensitive in nature, requiring security clearance for proper labelling by dedicated personnel leading to an insufficient developing process hindering the insertion of such technologies in this domain. Additionally, data may be specific to certain types of military sensors, such as infrared or multispectral sensors. When employing AI algorithms in military applications, it is essential that any recommendations and decisions made are compliant with safety and security regulations, considering the complex and continuously changing environment. Furthermore, an AI framework must be interpretable from the perspective of operators such as commanders.

Overall, data restrictions and singularity of the application affect severely the use of such models in a defence-oriented application. To this end, recent advancements in AI development, which specifically address these challenges, may be beneficial to overcome the aforementioned restrictions. The main principles of frugality, robustness, and explainability in AI have already proven to be auspicious solutions for these tasks. In this chapter, relevant topics and basic principles of such technologies are thoroughly analysed and presented acting as a proof of concept to strengthen their applicability in the defence domain. More specifically, an implementation approach is presented that involves all the aforementioned principles. The analysis is performed considering the specificity of the domain. A feasibility study is also presented to apply combinations of services in order to introduce higher semantically components and extract additional knowledge.

The rest of the chapter is organised as follows. The second provides insights into the unified approach to address the core challenge of few data availability for the defence domain while the third section presents the expected impact derived from its implementation. Finally, the fourth section concludes the chapter by describing the main results of this study.

Methodological Description

In general, artificial intelligence (AI) corresponds to the engineering and scientific methodologies that result in a system, which poses intelligent reasoning and behaviour. Recent technological advances have established AI as a technology with an impressive impact on almost every aspect of the modern socio-economic environment, thus significantly enhancing the capabilities of, among others, decision-making and support systems and autonomous processes, based on various types of data (text, visual content, video footage, etc.). State-of-the-art machine learning (ML) and deep learning (DL) techniques rely on mathematical methods that aim at extracting and exploiting application-relevant information from a large dataset. In this manner, AI systems are able to learn patterns and structures in an automated way by extracting correlations between complex data through the training/learning process. Hence, the availability of an amount of data comprises a major requirement to ensure a sufficient performance.

In a military context, data can be characterised as scarce or incomplete, sensitive (i.e. labelling can be performed only by experts with security clearance), and specific (i.e. data depend on specific types of military sensors, e.g. IR, multispectral, etc.) [5]. The recommendations and decisions to be proposed by AI algorithms in military applications must be acceptable, with respect to the safety rules and security regulations related to their uses and mission in a complex and rapidly changing environment. Moreover, from the operator’s point of view such as a commander, an AI framework must be acceptable with respect to its interpretability.

The described framework (Fig. 34.1) addresses the above challenges by incorporating novel approaches, algorithms, and tools as products of research to deliver some of the most prominent issues relevant to AI and increase their applicability and impact on defence systems. The framework relies on the following five pillars identified considering AI advancements and the peculiarity of the defence domain:

  • Data acquisition and improvements: Low availability of data can be addressed by applying data augmentation schemes while the timely process of annotation can be delivered with automated processes and innovative techniques.

  • Frugality in AI: To achieve high levels of accuracy when deploying an AI framework and exploiting it as an asset in modern warfare, operation-specific data must be available. Thus, the framework involves the application of dedicated models such as few-shot detections, domain adaptation, and transfer learning techniques.

  • Robustness in AI: Towards the same objective of reducing the size of the training dataset, principles such as discriminative, generative, and evolutionary learning will contribute to produce robust intelligence, surveillance, target acquisition, and reconnaissance (ISTAR) operations.

  • Explainability in AI: A comprehensive representation of the models’ outcomes is more than essential for an operator of a defence system. Hence, the appropriate measures and validation approaches are mandatory for quantifying their performance.

  • Improved situational awareness and mission planning: Additional knowledge can be extracted by ensuring a robust data exchange between the developed models. Hence, multimodal fusion schemes, threat assessment, and optimised mission planners can benefit from the application of AI models and deliver improved and comprehensive situational awareness.

Fig. 34.1
A schematic diagram of the A I-based framework in defense with 5 steps. It includes data acquisition and improvements, frugal A I, robustness in A I, explainability in A I, and A I for situation awareness, improved planning, and operational capabilities.

Schematic depiction of AI-based framework in defence

Full size image

Since the research is focused on defence applications, the first step will be to determine the user requirements of the defence domain which will drive the data needed to be processed. So, the defined end user requirements will then guide all the conducted research, designs, developments, and validations as this information is reflected in the required datasets. Data augmentation techniques incorporate generative learning principles [7] to produce additional data not only to increase the size of the available training set but also to reduce overfitting and improve the generalisation capabilities of the model. Random oversampling (ROS) comprises a naive approach which duplicates images randomly from the minority class until a desired class ratio is achieved. Intelligent oversampling techniques date back to synthetic minority over-sampling techniques (SMOTE) proposed by Chawla et al. [8]. SMOTE and its updated model, Borderline-SMOTE [9], develop new instances by interpolating new points from existing instances with the use of K-nearest neighbours. On the contrary, by their introduction from the research community, generative artificial nets (GANs) [5] displayed a clear superiority in producing synthetic data, mostly visual-related data. Architectures such as DCGAN [6] (Fig. 34.2), Progressively Growing GAN [10], CycleGAN [11], and Conditional GANs [12] seem to have the most application potential in the field. Similar progress has been observed in automating the process of annotation as it comprises a timely and resource-demanding (e.g. dedicated personnel) phase. To this end, deep learning architectures as proposed in [13, 14] can be rather beneficial.

Fig. 34.2
A schematic diagram of the D C G A N generator architecture. It includes the project and reshapes done by going through stride 1 of 1024, four stride 2s of 512, 256, 128, and 64 with convolutions 1 to 4 connecting them.

DCGAN, generator architecture [6]

Full size image

Frugal AI comprises the most crucial aspect of the presented framework. Frugal principles adopted in a detector model can contribute to reducing the necessity of large datasets. Considering also the lack of data for defence applications, the reuse of existing models adapted for these applications can be an efficient substitution (similar to the schema in Fig. 34.3). Hence, many AI models have been proposed which were tailored to support these applications like in [16, 17]. A similar rationale is used when the models incorporate transfer learning principles where the training process is not mandatory as pre-trained models are used for different applications. Indicatively, the models proposed in [18] display a sufficient performance, and by rectifying their behaviour, they could be utilised in the defence domain. Continual learning models [18] can contribute to higher performances with less data by continuously training the current version of a model when additional data are available. In more requirement-oriented models, environment adaptation [19] is another category of models that implements frugality and can be rather useful in defence applications.

Fig. 34.3
A flow diagram of frugal A I with photographs. The query with the image to be tested and support set examples are sent for match and feature fusion. It predicts and gives 3 analyses a, loss of spatial information, b, spatial misalignment, and c, uncertain class representations caused by average pooling.

Example of frugal AI for computer vision applications [15]

Full size image

Robust AI is most commonly envisaged with the application of hybrid AI principles. As is also implied by the used terminology, different types of learning are combined as well as the combination of data-driven learning with expert or domain knowledge. Thus, hybrid AI methods target to decrease the effects of dataset bias and overfitting. Various hybrid AI architectures have been proposed in numerous civilian applications [20]. For the military domain, concepts of integrating the advantages of different ML strategies with the ones provided by structure reasoning have been developed [21]. Additionally, the robustness of an AI model is commonly quantified with the use of dedicated metrics and the adoption of proper validations [22].

Explainable AI focuses mostly on interpreting appropriately and comprehensively a model’s detection outcome. For example, in computer vision, heatmap-based methods have been proposed that rely on saliency-like analysis [23]. Towards this objective, validation comprises a significant factor interpreting correctly the outcomes of a detection model whereas in defence applications considering the challenges, its significance is even larger [24].

Lastly, AI for improved situational awareness can lead in the upcoming years to the interest of the defence research community as its maturity level is still low compared to other domains. As in many applications, various sensors can be utilised having the same objective, and so, the collected data are fused to a higher integration layer. The most representative multimodal data fusion deep learning models from the perspectives of the model task, model framework, and evaluating data are based on deep belief networks [25], stacked autoencoders [26], etc. Another significant category of this pillar involves generative learning tools which account for enhanced threat assessment capabilities during mission definition and planning, incorporating cognitive analysis to provide quantitative and qualitative assessments of the potential threats such as in underwater mine warfare [27]. Finally, AI-based mission planning [28], a critical aspect in warfare, can complement the C2 functionalities towards delivering a complete decision support framework.

Expected Impact

All the aforementioned works can comprise a proper research base that is expected to produce advances in the development of new technologies which will be successfully applied in the context of AI for defence applications, covering all stages of a military operation supply chain. The framework and its core pillars described in the methodology section will contribute to the establishment of a strong AI-based defence industry enhancing the autonomy in the field of AI for military applications, leading to more efficient protection of critical infrastructures and military establishments and equipment. Furthermore, the frugal AI techniques will provide updates on the collection of data, contributing to a severe issue particularly in the defence domain while data-related restrictions can be overpassed.

Apart from the defence industry, deployment in a wide range of applications can be accelerated as the main attribute of the AI models is flexibility and efficient adaptation in unknown environments and different operating conditions. New business models and opportunities could emerge that will reduce the unit cost and competition growth among industries as one of the advantages of the models to be developed is that they will be built on the top of established knowledge and recent technology advances, ensuring effective services for potential adopters.

Beyond the utilisation in the defence domain, the present study will contribute to further exploring and understanding the usage of synthetic data and generative learning techniques. The data augmentation schemes that are studied are expected to create a pipeline of processing steps that can be used as guidance during the training phase of similar models. When applied to the input data, the robustness of the system could be increased and the overfitting of the model on the input data decreased. The methods introduced in the present study will be based on state-of-the-art data augmentation techniques that will explore further the utilisation of synthetic and cross-spectral data for deep learning. In addition, as far as annotation speed is concerned, the research will focus on increasing the annotation speed of tasks involving multimodal data by extending their usage to a production-level annotation pipeline of real-world datasets and using model architectures that capture inter-modality. The study will also deal with existing models’ knowledge and continuous learning in order to optimise the models that will acquire new knowledge without forgetting past knowledge when a new one is introduced. Focus will be given to the adaptation of few-shot learning techniques and selective prediction to increase the model’s self-awareness and aim to benefit the current state-of-the-art models.

Conclusions

In conclusion, the work presented in the current study aims to deliver new approaches, algorithms, and tools as research products based on the aforementioned pillars to address some of the most prominent issues relevant to AI applications and increase their applicability and impact on defence systems. Frugal AI models can account for operation-specific data and achieve high levels of accuracy when an AI framework is deployed in modern warfare. On the other hand, an operational-ready AI framework must remain reliable in case of natural or artificial perturbations and the guidelines of the framework must be trusted and understood to be acceptable by humans. Therefore, the developed models can focus on providing robustness and explainability.

Finally, the ultimate goal of the research is to contribute to the digitalisation of the defence domain by inserting novel and adjustable to current infrastructure systems in order to facilitate processes in the defence domain and exploit the benefits of human and machine cooperation.