Detecting attacks in network traffic is a challenge for data science: the data is
high-volume, labels are scarce, normal behavior drifts over
time, and attackers adapt to the detector. Static classifiers trained offline
struggle in this setting, because a defense is not a single prediction but a sequence of
decisions that trade detection accuracy against false alarms and operational cost.
In this talk, we frame security monitoring as a sequential decision problem and show how
Reinforcement Learning agents can learn adaptive detection and mitigation
policies directly from observations. We ground the discussion in Named Data
Networking, a future network architecture where content is requested by name rather than by
host address, and where new attacks emerge.
Keywords Reinforcement Learning, Network Security, Attack Detection, Named Data Networking,
Sequential Decision Making