An autonomous vehicle's LiDAR sensor scans the world several times per second, and almost nothing changes between two scans taken 100 ms apart: the road, the buildings, and even the moving cars are largely predictable from the previous sweep. Compression research exploits this temporal redundancy by predicting the current sweep from past sweeps and keeping only the "innovations" — the parts that could not be predicted. In parallel, security research has shown that a physically realizable adversary can hide an entire vehicle from a LiDAR-camera fusion detector by injecting as few as 20 fake points above the car's roofline, dropping detection recall from near 1.0 to below 0.2 on KITTI Exploring Adversarial Robustness of LiDAR-Camera Fusion Model in Autonomous Driving. Nobody has asked the natural question sitting between these two literatures: does the prediction-versus-innovation decomposition that temporal codecs compute also separate genuine geometry from attacker-injected geometry?
This proposal argues it can, but only partially, and that the partial answer is the point. The core mechanism is a temporal innovation gate: a multi-sweep predictor of the kind already used in neural LiDAR compression — MuSCLE's conditional entropy model over octree occupancy MuSCLE: Multi Sweep Compression of LiDAR using Deep Entropy Models — sits between the sensor and the detector. Every incoming point is scored by how well it is explained by the previous sweep under motion-compensated prediction. Real geometry, whether static background or a moving car, has temporal provenance; injected points optimized for a single frame do not. Points with high innovation and no motion-consistent explanation are flagged before detection. Crucially, we also build the attacker this defense invites: an adaptive adversary that optimizes its injected points across many sweeps so they are motion-consistent. The proposal's value is the two-sided characterization — how much robustness passive temporal redundancy buys, where it breaks, and what an explicit gate adds on top.
Core Claim
Single-frame physical point-injection attacks lose substantial potency when the victim performs multi-sweep temporal prediction, and an explicit innovation gate built on temporal redundancy modeling removes more of the attack than any static (non-temporal) purification at equal bitrate and compute — but a temporally adaptive attacker recovers a large fraction of the lost attack success, so temporal redundancy is a hardening mechanism, not a cure. This refines two existing findings rather than contradicting them. Zhu et al. showed that temporal cross-attention in BEVFormer creates a "safety buffer" (5.8 NDS points over non-temporal DETR3D under temporally universal patch attacks on nuScenes) while also showing that redundant BEV features admit catastrophic "ghost" activations (mAP collapsing from 37.0 to 0.2 under PGD) Understanding the Robustness of 3D Object Detection with Bird's-Eye-View Representations in Autonomous Driving — they diagnose the tension but propose no defense, and their analysis is camera-centric. SECVID showed that compression-plus-discretization scatters adversarial perturbations in 2D video Correction-based Defense Against Adversarial Video Attacks via Discretization-Enhanced Video Compressive Sensing — but its mechanism, breaking smooth pixel gradients, cannot apply to attacks that are already macro-geometric and discrete. Neither work touches 3D point streams or physically constrained point injection; nothing in the Lacuna file evaluates any 3D temporal codec under any physical attack, in either direction.
Research Program
The program has three parts. First, a measurement study: we take the Hiding Attack exactly as specified — one point per laser ray, discrete vertical emitter angles matching a Velodyne HDL-64E, narrow horizontal beam divergence Exploring Adversarial Robustness of LiDAR-Camera Fusion Model in Autonomous Driving — and replay it against a detector through four pre-processing pipelines: no codec, a classical codec (MPEG G-PCC quantization), a static deep codec (the same architecture with temporal context disabled), and a full multi-sweep temporal codec in the MuSCLE style MuSCLE: Multi Sweep Compression of LiDAR using Deep Entropy Models. This establishes whether temporal codecs purify incidentally. Second, the gate itself: we train a conditional occupancy predictor on clean multi-sweep sequences (the temporal-context machinery that already delivers up to 45% spatial bitrate reduction on SemanticKITTI in MuSCLE), and derive a per-point gate score
$$g(p_t) = r(p_t) \cdot \big(1 - \max_{q \in S_{t-1}} \kappa(p_t, q)\big),$$
where $r(p_t)$ is the occupancy-prediction residual for point $p_t$ at sweep $t$, and $\kappa$ measures how well $p_t$'s neighborhood is explained by motion-compensating the previous sweep $S_{t-1}$; a point is flagged when it is both unpredictable and unexplained. Third, the adaptive attacker: a multi-sweep Hiding Attack that optimizes point placement jointly across consecutive sweeps with ego-motion compensation, in the spirit of dynamic physical adaptation — ELA demonstrated that static physical attacks collapse in dynamic scenes (static AdvLB fell to 7–15% ASR while the embodied attack sustained up to 76%) Embodied Adversarial Attack: A Dynamic Robust Physical Attack in Autonomous Driving — but transferred to the LiDAR geometry channel and aimed at a temporal victim rather than a frame-wise classifier.
Proposed Experiments
Experiment 1 runs on KITTI Exploring Adversarial Robustness of LiDAR-Camera Fusion Model in Autonomous Driving against MVX-Net, matching the original setup: attack success rate at 20 and 200 injected points, and recall. The reference point is the published result — 39% ASR with 20 points, 99% with 200 — through no pre-processing. Each defense is run at matched bitrate and matched inference-time budget, and the attack is re-optimized white-box against each pipeline so no defense wins by obscurity. Experiment 2 evaluates the gate on SemanticKITTI SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences: clean-task cost is measured as downstream detection and segmentation accuracy on uncompressed and compressed streams, and robustness is measured against the single-frame and multi-sweep attacks. We also port the gate in front of BEVFormer on nuScenes nuScenes: A multimodal dataset for autonomous driving and re-run the 3D Consistent Patch Attack protocol of Zhu et al., where the reported comparison is BEVFormer outperforming DETR3D by 5.8 NDS points under temporally universal patches. Experiment 3 builds the adaptive multi-sweep attacker in CARLA CARLA: An Open Urban Driving Simulator, whose LiDAR sensor model and vehicle kinematics let us enforce physical emitter constraints across a moving ego-vehicle and target, then tests it against the gate, against adversarially trained detectors, and against BEVFormer's temporal attention. The decisive check is asymmetric by design: if the motion-consistent attacker fully restores ~99% ASR at 200 points against the gate, the honest conclusion is a negative result — passive and gated temporal redundancy both fail, and the paper's contribution becomes a precise characterization of why, in the same spirit as the 60-degree yaw wall that bounds patch attacks under perspective foreshortening Patch Attack Invariance: How Sensitive are Patch Attacks to 3D Pose?. Either outcome is publishable; the measurement is what is missing from the literature.
Baselines
The do-nothing reference is MVX-Net reading raw sweeps under the Hiding Attack, where the original paper reports 99% ASR with 200 points and recall falling from near 1.0 to below 0.2 on KITTI Exploring Adversarial Robustness of LiDAR-Camera Fusion Model in Autonomous Driving; any defense must be judged against these numbers, and a defense that leaves ASR above roughly 50% at 200 points has not changed the threat picture. The strongest existing competitors are, first, classical codec purification — MPEG G-PCC or Draco quantization applied before detection — because it is the cheapest possible "compression as defense" and directly inherits SECVID's hypothesis Correction-based Defense Against Adversarial Video Attacks via Discretization-Enhanced Video Compressive Sensing; it will be tuned to the same bitrate as the gate so the comparison isolates mechanism, not budget. Second, BEVFormer-style temporal cross-attention without any compression module, since Zhu et al. already demonstrate a 5.8-NDS temporal buffer under temporally universal patch attacks Understanding the Robustness of 3D Object Detection with Bird's-Eye-View Representations in Autonomous Driving; this is the strongest published temporal hardening and the gate must beat it or explain why it cannot. Third, adversarial training of the detector on injected-point augmentations, the standard robustness answer, run with the same GPU budget as gate training. The key ablation is our own architecture with the temporal context disabled — the static-vs-dynamic configuration that MuSCLE already separates for compression (temporal context alone reduces spatial bitrate by up to 45% on SemanticKITTI MuSCLE: Multi Sweep Compression of LiDAR using Deep Entropy Models) — run at matched parameters; if the static version purifies as well as the temporal one, the proposal's central ingredient is doing no work and the claim is falsified. Beating the baselines means, concretely: at 200 white-box-adapted points, ASR materially below the adversarially-trained and temporal-attention baselines at matched clean-task accuracy, with the clean accuracy cost of the gate under one point of detection mAP on uncompressed streams.
Why not just…?
Why not just quantize the coordinates, SECVID-style? Because the Hiding Attack is already discrete by construction: it satisfies one point per laser ray, discrete vertical emitter angles, and narrow beam divergence Exploring Adversarial Robustness of LiDAR-Camera Fusion Model in Autonomous Driving. SECVID's discretization works by breaking the smooth, high-frequency gradients that digital attacks depend on Correction-based Defense Against Adversarial Video Attacks via Discretization-Enhanced Video Compressive Sensing; a physically realizable point injection has no such gradients to break, and voxel-level quantization preserves the occupied structure the detector sees. Experiment 1 tests this directly, and we expect classical and static-deep codecs to leave ASR largely intact — which is itself an unstated and useful negative result for the "any compression purifies" intuition.
Why not just rely on the camera, since only LiDAR is attacked? The fusion literature already answers this: the Hiding Attack broke MVX-Net with a completely clean, unobstructed camera channel, reaching 39% ASR with 20 points Exploring Adversarial Robustness of LiDAR-Camera Fusion Model in Autonomous Driving, and Zhu et al. found that multimodal BEV fusion helps against natural corruptions but not against attacks on the joint feature space Understanding the Robustness of 3D Object Detection with Bird's-Eye-View Representations in Autonomous Driving. Dropping LiDAR also forfeits the metric accuracy that 3D detection exists for, so fallback-to-camera is not a comparable baseline on this question.
Why not just use temporal attention, which already works? BEVFormer's temporal cross-attention is the honest strongest competitor, and we run it as such. But the same paper that reported the 5.8-NDS buffer also reported that redundant BEV features admit ghost activations that collapse mAP from 37.0 to 0.2 under strong attack Understanding the Robustness of 3D Object Detection with Bird's-Eye-View Representations in Autonomous Driving: temporal attention accumulates history without deciding what in the current sweep is new, which is exactly the decision the gate makes. Whether that difference matters empirically is precisely what the matched comparison settles; if temporal attention matches the gate, we report that the cheap answer wins.
Why not just adversarially train the detector? This is a real threat and we do not argue around it: it is a trained baseline with an equal compute budget. Two things keep the proposal alive against it. Adversarial training defends against the perturbations in its training distribution, and our adaptive multi-sweep attacker (Experiment 3) is designed to sit outside the single-frame augmentation distribution that a naive training recipe would cover — if the trained detector holds against it, that is a strong positive finding for adversarial training and we will say so. And the gate is orthogonal to training: it can sit in front of a trained detector, so the relevant claim is marginal value, not replacement.
Decisive Checks
The proposal stands or falls on three numbers, all against white-box-adapted attacks at 200 points on KITTI: ASR through no codec (published as 99% Exploring Adversarial Robustness of LiDAR-Camera Fusion Model in Autonomous Driving), ASR through the gate, and ASR through the gate under the motion-consistent multi-sweep attacker. If the gate cuts ASR substantially below every matched baseline and the adaptive attacker recovers only part of it, temporal redundancy modeling is a defensible component of physical-world robustness. If the adaptive attacker erases the gap entirely, we will have mapped a physical limit of temporal defenses with the same clarity that the 60-degree pose wall was mapped for patch attacks Patch Attack Invariance: How Sensitive are Patch Attacks to 3D Pose? — and neither the compression literature (D-DPCC, patchDPCC, MuSCLE, VoxelContext-Net) nor the physical-attack literature (DTA, RAUCA, ELA, DynamicPAE) currently contains any measurement that distinguishes these outcomes.
References
- Exploring Adversarial Robustness of LiDAR-Camera Fusion Model in Autonomous Driving
- MuSCLE: Multi Sweep Compression of LiDAR using Deep Entropy Models
- Understanding the Robustness of 3D Object Detection with Bird's-Eye-View Representations in Autonomous Driving
- Correction-based Defense Against Adversarial Video Attacks via Discretization-Enhanced Video Compressive Sensing
- Embodied Laser Attack: Leveraging Scene Priors to Achieve Agent-based Robust Non-contact Attacks
- SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences
- nuScenes: A multimodal dataset for autonomous driving
- CARLA: An Open Urban Driving Simulator
- Patch Attack Invariance: How Sensitive are Patch Attacks to 3D Pose?