AutoDetect: Designing an autoencoder-based detection method for poisoning attacks on object detection applications in the military domain

Given the increasing sensor threat resulting from of application of more advanced sensors, proliferation of (low-cost) sensors and platforms (incl. drones) and application of automatic (AI) detection methods, solutions are sought for protection against detection and classification. Advancements in technology, materials and concepts have the potential of providing such protection. Furthermore, tactics and strategies can and must be altered to take current and future threats into account. As part of phase 1 of V2440: ‘Signature Dominance Land’ this report supplies an overview of the state-of-the-art with respect to materials, methods and technologies with (potential) application to signature management in the military domain.
Basically, two types of technology can be discerned. One the one hand, highly developed (high TRL) technology is available, often developed for other (civilian) purposes, such as display technology (e.g. LEDs, eInk, eChromes). In this case, the question is how to adapt this to the military domain. On the other hand, low TRL technologies are under development with high (but yet unknown) potential. Examples are metamaterials, smart-materials and 2-D material (e.g. graphene) with special properties, often tested under lab-conditions.
Specialized products developed for the military market (MOTS) present an increasing and widening market, with a rapid change in products. However, it remains difficult to assess the effectiveness of these products from the specifications and demonstration material provided (and stories found on internet). Therefore, there is an urgent need for (international) standards that allow for objective comparison of these products. For the moment, field and lab testing is required to evaluate these products.
International attention focuses around three main themes: i) multispectral camouflage, ii) adaptive camouflage and iii) countering (potential) threats posed by AI. The first means that solutions are sought that are able to counter different threat sensors operating in different spectral regions are the same time. Adaptability results in camouflage that better matches the environmental conditions and context and is therefore more effective. Recent and rapid developments in AI allow for automatic detection and classification, and the question is how to counter this threat.
It is also clear that a camouflage measure can only be effective when it is applied in an optimal way. Tactics form an important part of the art of remaining undetected, and these tactics have to be adjusted to the camouflage means available. Furthermore, one should also know the limitations and application area of a given camouflage measure or technology and be aware of the threats. This information (‘awareness’) can be used to derive an optimal strategy and optimal tactics for given operational conditions.