The Internet of Drones (IoD) enables coordinated
communication and collaboration among interconnected drones
to support diverse applications, including search and rescue,
disaster response, border surveillance, and military operations.
Restrained to the use open channels of communication and
limited resources, they are highly susceptible to diverse cyber-
attacks. One such assault is a collusion attack, where several
hacked drones work together to send fake sensor data and
confuse the system. We propose a federated learning-based
LSTM intrusion detection framework to tackle this vulnerability,
which probes sliding windows of drone data and learns patterns
over time. We conducted experiments on simulated search and
rescue missions with different attack scenarios. We compared the
proffered frameworks with preset-threshold intrusion prevention
systems and machine learning models. The proposed framework
performed perceptively better than pre-configured intrusion
prevention systems and found suitable for deployment on fog
nodes. Over five independent experimental runs, our centralized
LSTM model achieves an F1-score of 0.963 (AUC = 0.998) and
rapidly isolating compromised nodes within 1.46 timeslots. Our
federated framework achieves a global classification F1-score of
0.717 (AUC = 0.956), and demonstrating it is suitability for fog-
node deployment. We discuss safety-critical trade-offs concerning
false honest-node isolation rates and outline future directions
to bridge the centralized-federated utility gap while integrating
robust mutual authentication protocols.

Collusion attack detection,federated learning,fog computing,Internet of Drones,intrusion prevention system,Long Short-Term Memory.