Perturbation Strategies
Overview
Adversarial perturbations are small, carefully crafted modifications to LiDAR point clouds that degrade SLAM performance while remaining imperceptible. This document describes the 17-parameter genome encoding and the effectiveness of each perturbation strategy against MOLA SLAM.
The perturbation generator implements state-of-the-art adversarial techniques based on recent research:
- FLAT: Flux-Aware Imperceptible Adversarial Attacks (ECCV 2024)
- SLACK: Attacking LiDAR-based SLAM (arXiv 2024)
- ICP Attack: Adversarial attacks on ICP registration (arXiv 2403.05666)
- ASP: Attribution-based Scanline Perturbation (IEEE 2024)
Attack Objectives
An effective adversarial perturbation must balance two goals:
- Maximize localization error - Cause MOLA to produce inaccurate trajectory estimates
- Minimize detectability - Keep perturbations small enough to avoid detection
The NSGA-III optimization finds perturbations that achieve the best trade-off between these objectives.
17-Parameter Genome
The genome encodes 17 continuous parameters in the range [-1, 1], which are then scaled to their respective physical ranges. Each parameter controls a different aspect of the perturbation.
Basic Perturbations (Parameters 0-6)
Parameters 0-2: Noise Direction (3D Vector)
A directional bias applied to all points.
- Parameter 0: X component of noise direction
- Parameter 1: Y component of noise direction
- Parameter 2: Z component of noise direction
The direction vector is normalized and scaled by noise intensity. This creates a systematic drift in a specific direction rather than random noise.
Why it works: Directional bias accumulates over time, causing consistent odometry drift that SLAM cannot correct without loop closures.
Parameter 3: Noise Intensity
Controls the magnitude of Gaussian noise added to point coordinates.
- Range: 0-5cm standard deviation
- ATE correlation: +0.370 (strong positive effect)
- Perturbation correlation: +0.366 (increases detectability)
Why it works: Random noise degrades ICP point-to-plane alignment accuracy. Higher noise means less precise feature matching between consecutive frames.
Parameter 4: Curvature Targeting
Targets high-curvature regions (edges, corners) with stronger perturbations.
- Range: 0-100% strength
- ATE correlation: +0.227 (moderate positive effect)
- Perturbation correlation: -0.025 (minimal impact on detectability)
Why it works: SLAM systems rely on geometric features (edges, corners, planes) for alignment. Corrupting these features specifically degrades matching quality.
Parameter 5: Dropout Rate
Percentage of points randomly removed from each frame.
- Range: 0-30%
- ATE correlation: +0.055 (weak positive effect)
- Perturbation correlation: +0.105 (low detectability impact)
Why it works: Removing points reduces feature density, degrading ICP convergence and loop closure detection. However, MOLA is relatively robust to moderate dropout.
Parameter 6: Ghost Point Ratio
Adds synthetic points near real measurements.
- Range: 0-10% additional points
- ATE correlation: -0.086 (slightly negative effect)
- Perturbation correlation: +0.110 (increases detectability)
Why it works: Ghost points create false correspondences in ICP. However, MOLA's outlier rejection often filters these, making this less effective.
Cluster Perturbations (Parameters 7-10)
Parameters 7-9: Cluster Direction (3D Vector)
Direction for cluster-based perturbations.
- Parameter 7: X component (ATE correlation: -0.261)
- Parameter 8: Y component (ATE correlation: +0.003)
- Parameter 9: Z component (ATE correlation: +0.036)
Parameter 10: Cluster Strength
Intensity of cluster-based perturbations.
- Range: 0-100%
- ATE correlation: +0.071 (weak positive effect)
- Perturbation correlation: -0.037 (minimal impact)
Why it works: Clusters of points are shifted together, maintaining local structure while introducing global distortion.
Advanced Perturbations (Parameters 11-16)
Parameter 11: Spatial Correlation
Controls how perturbations are correlated spatially.
- Range: 0-100%
- ATE correlation: -0.167 (negative effect)
- Perturbation correlation: -0.070 (reduces detectability)
Why it works: Spatially correlated perturbations appear more natural than random noise. However, high correlation may actually help SLAM by preserving local structure.
Parameter 12: Geometric Distortion (ICP Attack)
Applies systematic geometric distortions (scaling, shearing).
- Range: 0-100%
- ATE correlation: +0.357 (strong positive effect)
- Perturbation correlation: +0.970 (very high detectability!)
Why it works: ICP assumes rigid transformations. Systematic distortions violate this assumption, causing alignment failures. However, this is highly detectable due to large point displacements.
Parameter 13: Edge Attack (SLACK-inspired)
Targets edge and corner points with perpendicular shifts.
- Range: 0-100%
- ATE correlation: +0.033 (weak positive effect)
- Perturbation correlation: -0.149 (reduces detectability)
Why it works: Shifting edge points perpendicular to their principal direction maximizes ICP confusion while minimizing visible distortion.
Parameter 14: Temporal Drift
Accumulating bias across frames over time.
- Range: 0-100%
- ATE correlation: +0.627 (strongest positive effect!)
- Perturbation correlation: +0.237 (moderate detectability)
Why it works: Temporal drift is the most effective attack because: - Bias accumulates consistently over the trajectory - Prevents loop closure detection (locations appear different over time) - SLAM cannot correct accumulated drift without recognizing revisited places
Parameter 15: Scanline Perturbation (ASP-inspired)
Perturbs points along their laser beam directions.
- Range: 0-100%
- ATE correlation: +0.595 (second strongest effect!)
- Perturbation correlation: +0.392 (moderate-high detectability)
Why it works: Perturbations along scanlines simulate realistic sensor interference (dust, particles). This systematically affects range measurements in a physically plausible way.
Parameter 16: Strategic Ghost Placement
Places ghost points near geometric features.
- Range: 0-100% (activates when > 50%)
- ATE correlation: +0.344 (strong positive effect)
- Perturbation correlation: -0.026 (minimal detectability impact)
Why it works: Ghost points placed near features create ambiguous correspondences that ICP cannot easily reject, unlike random ghost points.
Effectiveness Rankings
Based on correlation analysis from genome12 results (347 valid evaluations):
Most Effective for Increasing ATE
| Rank | Parameter | ATE Correlation | Why Effective |
|---|---|---|---|
| 1 | Temporal Drift | +0.627 | Accumulates over time, breaks loop closure |
| 2 | Scanline Perturbation | +0.595 | Systematic range errors, physically realistic |
| 3 | Noise Intensity | +0.370 | Degrades ICP alignment precision |
| 4 | Geometric Distortion | +0.357 | Violates ICP rigid transformation assumption |
| 5 | Strategic Ghost | +0.344 | Creates ambiguous feature correspondences |
Defending Against Perturbations
Understanding these attacks informs defense strategies:
- Temporal consistency checking: Detect sudden changes in point cloud statistics between frames
- Multi-sensor fusion: Combine LiDAR with camera or IMU for redundancy
- Learned anomaly detection: Train classifiers to detect adversarial patterns
- Robust loop closure: Use multiple verification methods for place recognition
- Range verification: Cross-check LiDAR ranges with expected environment geometry