KTH's LASE-AD Catches Lying Drone Sensors Before They Cause a Crash
A new framework called LASE-AD keeps control systems running when attackers spoof their sensors, using Bayesian inference and active probing to isolate compromised inputs.
ProTakeaways
- KTH researchers introduce LASE-AD, a framework bridging sensor attack detection and recovery in cyber-physical systems.
- Perception pipelines are modeled as bipartite graphs, turning anomaly alerts into a Bayesian network over sensor attack states.
- Active probing injects control inputs that maximize KL divergence between attack hypotheses to speed up identification.
- A threshold-based probing policy is justified via a simplified POMDP under Blackwell-dominance conditions on sensor informativeness.
- On extended CartPole, LASE-AD keeps the pole upright while all baselines fail 100% under prolonged attacks.
- No public code yet; real-world UAV and ground-vehicle validation is flagged as future work.
When someone spoofs the GPS on a drone, jams a gyroscope with acoustic noise, or injects bias into a wheel encoder, most control systems have two bad options: ignore the alarm and crash, or discard all sensor data and coast on stale predictions until something goes wrong. A new paper from KTH Royal Institute of Technology proposes a third option, called LASE-AD, that figures out which specific sensors are lying and quietly drops them from the state estimator while keeping the rest of the system running.
The work, Active Bayesian Inference for Robust Control, targets a real gap in cyber-physical systems research. Detection has received plenty of attention, but detection alone is insufficient: if the controller cannot respond appropriately, the system remains vulnerable. Existing recovery approaches either swap in a conservative backup controller or use outlier-robust Kalman filters that downweight suspicious measurements without ever identifying which sensor is the culprit.
Turning a perception pipeline into a Bayesian network
The core insight is structural. Modern autopilots like PX4 fuse heterogeneous sensors such as cameras, IMUs, encoders, and GPS through a preprocessing pipeline before anything reaches the Kalman filter. The authors model that pipeline as a bipartite graph mapping sensors to the components of the state estimate they influence. Combined with alerts from an off-the-shelf anomaly detector, the graph becomes a Bayesian network whose latent variables are the attack states of individual sensors.