TACTIC: Temporal and Context-Aware LLM Tactical Planning for Roadside LiDAR Attacks
Yiming Gao, Shaocheng Luo
Embedded Files
Overview
Overview
How does surrounding traffic change the impact of a LiDAR attack? TACTIC studies this question through scene-dependent tactical planning in CARLA. It combines roadside LiDAR measurements and camera imagery with multimodal language-model reasoning to represent relationships among vehicles and evaluate context-sensitive failure modes.
Open Science Policy
Open Science Policy
To comply with the open science policy, we will ensure that our source code and datasets for simulations are publicly available on GitHub.
Contributions
Contributions
We formulate physical LiDAR attack execution as a scene-dependent tactical planning problem, focusing on dynamic vehicle-to-vehicle interactions whose feasibility depends on evolving relational traffic context.
We develop TACTIC, which combines object-level local perception and roadside imagery with multimodal LLM reasoning to construct semantic scene graphs and select complementary physical attack primitives and parameters.
We ground LLM tactical decisions with measured vehicle states and calibrated physical constraints, and introduce an asynchronous perception–LLM architecture for high-rate scene-change detection and replanning. Experiments demonstrate higher collision success than fixed and random policies and faster adaptation to changing scenes.
System overview: roadside sensing, semantic scene reasoning, feasibility checks, and asynchronous replanning.
Demo Videos
Demo Videos
case_1_rear_end.mp4Experiment 1 — Push-Away (Rear-End)
Experiment 1 — Push-Away (Rear-End)
Point-cloud push-away shifts the perceived range of the lead vehicle; the target's car-following controller closes the gap and collides.
case_1_emergency_brake.mp4Experiment 2 — Phantom-Object Braking
Experiment 2 — Phantom-Object Braking
An 8 m phantom wall is injected into the target's perceived scene, triggering emergency braking.
CASE_2.mp4Experiment 3 — Scene Switching I
Experiment 3 — Scene Switching I
Traffic topology changes mid-episode; high-rate local perception detects the mutation, and the agent replans with the new scene graph.
CASE_3.mp4Experiment 4 — Scene Switching II
Experiment 4 — Scene Switching II
A second switching scenario demonstrating cached-topology reuse and asynchronous replanning under continuing MLLM latency.
Key Results
Key Results
The experiment runs 280 primary randomized CARLA trials across feasibility, policy, perception, timing, and benign-control evaluations. The policy comparison below uses 20 trials per group.
Decision-Source Comparison:
TACTIC full policy: 100% collision rate (20/20).
Restricted LLM: 75% collision rate (15/20).
Random selection: 60% collision rate (12/20).
Fixed rule: 35% collision rate (7/20).
Perception-Input Ablation:
Joint physical measurements and imagery: 100% collision rate (20/20).
Physical-only: 65% collision rate (13/20).
Image-only input: 75% collision rate (15/20).
Faster scene updates: asynchronous Δ refresh reduced scene-mutation response latency from 7.4 s to 2.0 s.
Benign control: no collisions occurred in 20 attack-free trials. These outcomes describe the evaluated simulation settings.
Page updated
Report abuse