This paper studies the network isolation attack, a devastating type of attacks on cyber-physical systems. In this attack, an adversary compromises a set of nodes that enclose a region in order to isolate the region from the rest of the network. Assuming that the compromised nodes wish not to be detected, we propose a solution to defend against the network isolation attack. Our goal is to achieve the following security guarantee: either a legitimate node can successfully deliver a message to another legitimate node, or the network control center can identify a small set of suspect nodes, which are guaranteed to contain a compromised node. Toward achieving this goal, we develop two protocols: one is for secure delivery of messages among nodes and the other is for secure collection of messages from nodes at the network control center. We show that our proposed protocols are provably secure, i.e., attain the aforementioned security guarantee. Further, our protocols achieve this guarantee with overhead that is orders-of-magnitude smaller than existing baseline protocols. Our proposed protocols are thus scalable for large networks.