We study efficient and lightweight Intrusion Detection Systems (IDS) for ad-hoc networks via the prism of IPv6-enabled Wireless Sensor Actuator Networks. These networks consist of highly constrained devices able to communicate wirelessly in an ad-hoc fashion, thus following mesh networks. Current state-of-the-art (IDS) have been developed taking into consideration regular computer networks, and as such they do not efficiently addresses the paradigm of ad-hoc networks. In this work we firstly identify a trade-off between the communication and energy overheads of an IDS (as captured by the number of active IDS agents in the network) and the performance of the system in terms of successfully identifying attacks. In order to fine tune this trade-off, we model such networks as Random Geometric Graphs; a rigorous approach that allows us to capture underlying structural properties of the network. We then introduce a novel IDS architectural approach by having only a subset of the nodes acting as IDS agents. These nodes are able to efficiently detect attacks at the networking layer in a collaborative manner by monitoring locally available network information provided by IoT routing protocols such as RPL. Our detailed experimental evaluation demonstrates significant performance gains in terms of communication overhead and energy dissipation while maintaining high detection rates. We also show that the performance of our IDS in ad-hoc networks does not rely on the size of the network but on fundamental underling network properties, such as the network topology and the average degree of the nodes.