Program statusOn track
Thesis research program
Dendroaspis: Efficient On-Device Behavioral Anomaly Detection on Kernel Telemetry with Selective State-Space Models
A distributed behavioral anomaly-detection study in which lightweight host-level state-space models learn normal endpoint behavior and escalate only suspicious sequences for broader reasoning.
This portal tracks the current plan, weekly progress, architecture decisions, risks, and experiments for the thesis program.
Immediate priority
Telemetry Collection Design
Finalize and validate the thesis telemetry collection contract across Tracks A, B, and C before experiment entities are provisioned and baseline collection begins.
Open current sprintResearch agenda
Research questions
Which self-supervised objective best captures host-behavioral anomalies on a causal, on-device backbone?
What is the minimal, transferable feature encoding, and which discriminative axes carry the attack signal?
Does fusing independent discriminative axes improve detection and raise the mimicry floor?
How far does a host-trained agent transfer to a different host?