Targeted Trajectory Redirection in Sequential and Tracking Systems

36 concepts23 papers

Traditional adversarial attacks are chaotic. They add pixel noise to make an image classifier output "toaster" instead of "guacamole," or they disrupt an object detector so a vehicle completely vanishes for a single frame. In autonomous systems, however, these "untargeted" failures are often harmless. Because modern tracking and navigation pipelines are built to filter out transient noise, a single anomalous frame is simply ignored or smoothed over by temporal filters.

To truly compromise these systems, an adversary cannot just cause random blindness; they must execute targeted trajectory redirection. Instead of breaking the system, the attacker subtly bends its perception over time, forcing the model to track or predict a specific, completely fabricated path. By understanding why sequential models trust history, attackers can turn temporal consistency into a vulnerability.

Weaponizing Temporal Consistency

Why does targeted redirection work against state-of-the-art trackers and forecasting models? The secret lies in how these models maintain internal state.

Consider a Multiple Object Tracking (MOT) system. It associates detections across video frames by checking if a new bounding box matches the expected position based on previous frames. If an attacker injects a massive spatial jump, the tracker rejects it as an anomaly. However, as Yantao Lu et al. demonstrate in Fooling Detection Alone is Not Enough: Adversarial Attack against Multiple Object Tracking, if an adversary introduces tiny, incremental translations to the bounding boxes over just a few frames, these changes are small enough to satisfy data association constraints. Crucially, they alter the tracker's internal velocity estimates. Once the internal velocity is manipulated, the tracker willingly steers the object's perceived trajectory along the attacker's desired path.

[Frame t]     True Position -> Small Perturbation -> Tracker Accepts Box
                                                          ↓
[Frame t+1]   Updated Velocity Estimate Based on Perturbed Box
                                                          ↓
[Frame t+2]   Tracker Predicts New False Position, Expecting Smooth Motion

This exploitation of sequential history extends to autoregressive time-series and reinforcement learning models. In these systems, each prediction or action is conditioned on a window of past observations. Yiding Chen et al. frame this as an optimal control problem in Optimal Attack against Autoregressive Models by Manipulating the Environment. By planning perturbations over a multi-step horizon, an attacker can use minimal energy to nudge the environment state, causing the autoregressive model to recursively hallucinate a target trajectory.

Similarly, Xiaolin Sun et al. show in Diffusion Guided Adversarial State Perturbations in Reinforcement Learning that conditioning each adversarial observation on previously perturbed inputs creates a self-consistent, fictional history. The victim agent follows this false trail into a failure state because the sensor data remains internally logical over time, bypassing standard anomaly monitors.

From Prediction to Planning Error Propagation

The ultimate danger of trajectory redirection is not that an internal variable changes, but that this error propagates directly into physical actions. In autonomous driving pipelines, data moves from perception to trajectory prediction, and finally to motion planning.

As Marsalis T. Gibson et al. explain in Hacking predictors means hacking cars: Using sensitivity analysis to identify trajectory prediction vulnerabilities for autonomous driving security, trajectory prediction models are highly sensitive to minor input variations. Because the downstream planning module treats these predictions as ground-truth forecasts of the environment, compromising the predictor is functionally equivalent to hijacking the physical vehicle's behavior.

Instead of optimizing for raw displacement error, an attacker can optimize for high-level semantic goals. Qingzhao Zhang et al. demonstrate in On Adversarial Robustness of Trajectory Prediction for Autonomous Vehicles that an adversary can manipulate predictions to simulate a sudden lane change or braking event by a neighboring vehicle. The downstream planning module, sensing an imminent collision, is forced to execute extreme emergency maneuvers like hard braking.

To make these threats stick against safety filters, Yi Yu et al. developed a multi-frame framework in Enduring, Efficient and Robust Trajectory Prediction Attack in Autonomous Driving via Optimization-Driven Multi-Frame Perturbation Framework that generates persistent, multi-frame perturbations. This ensures the false threat isn't dismissed as a transient glitch, successfully inducing "ghost braking" or collisions in simulated environments.

Remarkably, this redirection can even be achieved without digital pixel manipulation. Adrien Gauffriau et al. show in Autonomous Drone Interception with Reinforcement Learning that an attacker drone can use its own physical positioning to exploit a target drone's anti-collision safety protocols. By maintaining a specific relative position, the attacker forces the target into continuous avoidance maneuvers, effectively "herding" it along a malicious trajectory without ever touching its code.

The Optimization Mechanics of the False Path

To force a model along a precise decoy path, attackers must solve a complex optimization problem across multiple time steps. Simple gradient descent on a single frame is insufficient because tracking benchmarks often feature automatic re-initialization mechanisms that reset the tracker if it drifts too abruptly, as noted by Siyuan Liang et al. in Efficient Adversarial Attacks for Visual Object Tracking.

To overcome this, online and incremental optimization strategies are used. In SPARK: Spatial-Aware Online Incremental Attack Against Visual Tracking, Qing Guo et al. formulate the attack to maximize the tracker's response specifically at the coordinates of a pre-determined decoy path, updating the perturbation incrementally frame-by-frame to maintain temporal consistency.

Loss = - TrackerResponse(Decoy_Coordinates) + Smoothness_Penalty(Perturbation)

In the physical world, this optimization translates to spatial coordinates or environmental modifications:

By shifting the adversarial paradigm from disruption to direction, these methods reveal that the temporal consistency mechanisms designed to make autonomous systems robust can be the very levers used to control them.

Key Foundational Work

SPARK: Spatial-Aware Online Incremental Attack Against Visual Tracking (2020) — contributed 2 concept elements to this direction.
Efficient Adversarial Attacks for Visual Object Tracking (2020) — contributed 2 concept elements to this direction.
Tactics of Adversarial Attack on Deep Reinforcement Learning Agents (2017) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_targeted_trajectory_redirection_c685b31b,
  title = {targeted trajectory redirection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/targeted-trajectory-redirection/rdr_c685b31bc37545528f3af40385ea1a07},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_targeted_trajectory_redirection_c685b31b_v1,
  title = {targeted trajectory redirection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/targeted-trajectory-redirection/rdr_c685b31bc37545528f3af40385ea1a07/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake