Connected vehicles increasingly share what their sensors see, so that a car approaching a blind corner can "see through" the building using a report from a vehicle or roadside camera that has line of sight. The catch is that the receiving car cannot check a claim about space it cannot see. A compromised sender can report a fake pedestrian or delete a real truck in exactly those hidden regions. Current defenses compare the claim against the receiver's own view or against other senders. Both fail when the claim is about the receiver's blind spot and only one sender can see it. A published attack, Pretend Benign, exploits this and reports near-100% stealth against consensus defenses.
We propose that hidden-region claims are verifiable even though they cannot be observed directly, because a real object leaves physical traces in space the receiver can see. A car hidden behind a building now had to come from somewhere. Over the last few seconds, the receiver saw many regions empty, and a real object's past path must avoid all of them. Visible drivers near the hidden object, including the sender itself, should react to it. The sender must also have had an unobstructed line of sight to what it reports. We turn these consequences into a sequential test that runs as the receiver drives. The test outputs a probability that each claim is real, and the planner treats claims according to that probability rather than accepting or rejecting them outright.
Core Claim
The main mechanism is a free-space history test, and it is where this work departs from the static-world habits of both security and planning. Let $F_\tau$ be the set of points the ego vehicle observed as empty at past time $\tau$. An ego-observed empty region is only known to be empty at that instant, so we do not merge these sets into a static map. A claim $c$ reports a hidden object with state $x_c$ (position, heading, velocity, size) at time $t$. The claim is physically admissible only if some dynamically feasible past path reaches $x_c$ while avoiding every region the ego saw empty:
$$\exists\, x(\cdot) \in \mathcal{R}_{[t-T,t]}(x_c) \;\text{ such that }\; x(\tau) \notin F_\tau \;\; \text{for all } \tau \in [t-T, t],$$
Here $\mathcal{R}_{[t-T,t]}(x_c)$ is the set of backward trajectories of length $T$ that end at $x_c$ and respect vehicle or pedestrian limits on speed, acceleration and turning. Checking this is a backward reachability computation over the ego vehicle's own time-indexed occupancy record, so it needs no second witness.
Softer evidence is combined with it as a log-likelihood ratio. The ratio compares "claim real" against "claim fabricated" and draws on three sources: the observed motion of visible witnesses $w$, including the sender; ray-casting against the HD map to ask whether the sender's claimed pose could have seen $x_c$; and, eventually, direct observation once the object leaves the shadow:
$$\Lambda_t(c) = \sum_{w} \log \frac{p(\text{motion}_w \mid c \text{ real})}{p(\text{motion}_w \mid c \text{ fake})} + \log \frac{p(\text{sender visibility} \mid c \text{ real})}{p(\text{sender visibility} \mid c \text{ fake})}.$$
The key empirical claim is about the attacker, not the defender. Adding the free-space history test sharply shrinks the region where a fabricated object can sit undetected. The region shrinks from "all occluded space", which is what Pretend Benign exploits, to "occluded space never observed empty and reachable from never-observed space". The second region collapses quickly as the ego approaches an intersection and sweeps its sensors across the approach lanes. We will also measure the region that remains unverifiable and report it honestly. The planner hedges only against this residual set. This replaces the uniform worst-case assumption that causes over-braking in Occlusion-Aware Risk Assessment.
How This Relates to Prior Work
Three lines of work motivate the gap. CP-Guard finds malicious senders by recursively splitting groups and comparing each group's bird's-eye-view (top-down grid) segmentation against the ego's own. The Latent Guardian (Cerberus) checks whether shared feature maps look natural, using structural similarity, cosine distance and activation energy. Cerberus also loosens its thresholds when the ego's view is uncertain. Both tools judge single snapshots against what the ego sees. Neither has any basis for judging a claim located entirely in the ego's shadow, and Cerberus's loosening makes that region easier to attack. Pretend Benign confirms the gap: it concentrates perturbations in the "imperceptible region" and proposes no defense.
Pretend Benign also constrains the design, and we treat this as a negative result. It uses a spatiotemporal transformer with motion prediction to make its fake signals consistent across frames. A defense that only checks whether a ghost moves plausibly is therefore likely to be evaded. For this reason we do not rest the proposal on kinematic plausibility alone.
Two classes of prior work supply the methods we reuse. Control-theoretic residual checking, framed in Research Challenges for the Security of Control Systems and developed in Secure Safety Filter and Combating False Reports, compares physics predictions against sensor reports. It applies only to the host system's own state. We extend the same idea to external objects the host cannot sense.
Reachability over unobserved space in Safe Motion Planning in Unknown Environments is our backward-reachability tool. We repurpose it from bounding the ego's own risk to bounding what an attacker can fake. Behavior Planning with Social Perception infers hidden hazards from human drivers' braking using inverse reinforcement learning, which learns a driver's cost function from observed behavior. That work assumes honest drivers and no radio messages. We invert its use: witness behavior becomes evidence about a digital claim. Because this evidence is fragile to erratic drivers, we make it secondary to the free-space test.
Collaborative perception systems are what we protect. SparseCoop sends explicit per-object kinematic states, which makes our test cheap to apply but is itself unverified. CoDynTrust down-weights stale features under latency, but it trusts fresh, confident, fabricated features. Planners that depart from the static world, such as BeTopNet, assume every reported agent exists. We test how much a single ghost disrupts them.
Research Program
We will build four components.
The first is a time-indexed free-space memory for the ego vehicle. It stores LiDAR-observed empty cells with timestamps over a horizon $T$ of about 3 to 5 seconds. It deliberately does not merge observations into a static occupancy map, because a static map would wrongly rule out objects that entered a region after it was seen.
The second is a backward-reachability checker for each claimed object class. It uses conservative bounds on speed and acceleration, uses map lane topology for vehicles, and is computed on a coarse grid so it runs within one perception cycle. It works on object-level claims directly, as in SparseCoop. For feature-map systems, it works on detections decoded from the fused feature map with and without each sender's contribution.
The third is the witness and visibility model. The sender's reported pose is checked against the ego's own detection of the sender when the sender is visible. Line of sight is ray-cast against map buildings. Witness reactions are scored with a learned conditional motion model trained on clean data: a standard motion forecaster conditioned on the presence or absence of the hidden object.
The fourth is a planner interface. Each hidden object enters the planner weighted by its posterior probability. Phantom-risk particles, in the style of Occlusion-Aware Risk Assessment, are placed only in the residual unverifiable set.
On the attack side, we will reimplement Pretend Benign. We will also build a strictly stronger adaptive attacker that knows the ego's pose history and the map. This attacker places ghosts only where the free-space test cannot refute them and synthesizes witness-consistent sender motion wherever the sender is under its control. The analysis target is the size of the residual unverifiable region as a function of distance to the conflict point. This size tells us whether the method buys useful lead time before the ego sees the region itself.
Baselines
The naive reference points are no fusion and undefended fusion. For no fusion, OPV2V reports PIXOR at [email protected] of 0.406 in the default towns. For clean Attentive Intermediate Fusion, the same paper reports 0.734. Undefended fusion under attack sets the harm level, and no fusion sets the safety floor for any defense that simply rejects everything hidden.
The strongest existing detector-level defense is Cerberus. It reports an attack success rate of 0.05% and AP restored to 0.88 on OPV2V against the attacks it was evaluated on. However, no reported result was found on Lacuna for Cerberus under the Pretend Benign attack. We will run Cerberus with its published thresholds and also with thresholds re-tuned on our validation attacks, so that its tuning budget matches ours. Beating it means a substantially lower attack success rate on ghosts placed in the imperceptible region under Pretend Benign and under the adaptive attacker, while matching its AP on clean data and on attacks in visible regions.
The consensus baseline is CP-Guard. It reports mIoU recovered to 39.34% under PGD attack on V2X-Sim, against 40.45% clean and 14.34% undefended. We run it on V2X-Sim with identical attack budgets. Pretend Benign already reports near-100% stealth against ROBOSAC and feature-level defenses, so we expect consensus to fail in the imperceptible region. We include CP-Guard to confirm this on our own splits rather than assume it.
The planning baseline is phantom-particle risk from Occlusion-Aware Risk Assessment, which reports a 95th-percentile collision rate dropping from 12.64% to 2.61% on its Ann Arbor intersections. We re-create comparable intersections in CARLA/OpenCDA and compare collision rate and progress at matched settings.
Three ablations isolate the ingredients. The first uses kinematic plausibility only, checking whether the claimed track obeys motion limits over time. The second uses free-space history only. The third uses witness evidence only. Beating kinematics-only under the adaptive attacker is the specific result that justifies the free-space memory.
Why not just…?
Why not just check that the claimed object's motion is physically possible across frames? This is the obvious extension of residual checking to external objects, and we run it as an ablation. Pretend Benign already generates temporally consistent, motion-predicted fake signals, so an attacker can produce smooth, feasible ghost tracks at little cost. A plausible track can still pass through space the ego saw empty one second earlier. The free-space test catches that; a plausibility check does not. If kinematics-only matches the full method under the adaptive attacker, the free-space memory adds nothing and the claim fails.
Why not just treat every occlusion as dangerous and ignore shared data there? This conservative planner is a real competitor for safety, and it is our safety floor. Its cost is the braking at every blind corner that Occlusion-Aware Risk Assessment itself documents. It also discards the occluded-object gains reported by V2VNet, over 60% improvement on objects with few or no LiDAR points. The honest threat is timing. If our test settles a claim only when the ego is about to see the region anyway, verification adds nothing over waiting. We measure decision lead time directly to settle this.
Why not just use sender reputation or tighter feature-space thresholds? Reputation fails against insiders, and it fails against Pretend Benign by construction, since that attack stays consistent with consensus in visible regions. Tightening Cerberus's thresholds in occluded regions should reduce attacks but also raise false rejection of real hidden objects. This is the trade-off Cerberus loosened its thresholds to avoid. We plot both methods on the same curve of attack success against false rejection of real hidden objects to show whether physical evidence moves the curve or merely picks a different point on it.
Proposed Experiments
Open-loop detection experiments use OPV2V and V2XSet sequences, since multi-frame scenes are needed for free-space memory, and V2X-Seq for real-world sequential tracking. We label every ground-truth object as visible, partially visible or fully hidden from the ego. We inject Pretend Benign attacks and adaptive-attacker ghosts and deletions only into fully hidden regions. Metrics are:
- attack success rate in hidden regions;
- AUROC separating real from fabricated hidden objects;
- false rejection rate on real hidden objects, which is the cost;
- clean [email protected];
- decision lead time, meaning seconds between the verdict and the ego's first direct observation.
For deletions, where a sender suppresses a real object, the same free-space test runs in reverse. An object seen entering the shadow cannot silently vanish. This is a secondary check.
Closed-loop experiments run in CARLA/OpenCDA intersections modeled on the Ann Arbor layouts, with honest and malicious senders and human-model bystanders. Metrics are collision rate, ghost-induced hard brakes, time to clear the intersection, and jerk. We compare against undefended fusion, Cerberus-filtered fusion, phantom particles, and our posterior-weighted planner. We also measure how often BeTopNet-style interactive prediction is disrupted by a single accepted ghost.
Two checks are decisive. First, under the adaptive attacker, the residual unverifiable region must shrink enough that the full method beats kinematics-only by a clear margin in hidden-region attack success, at equal false rejection. Second, in closed loop, the method must match phantom particles on collision rate while clearing intersections faster, with lead time well above zero. If the residual region stays large until the ego can see the region itself, we will report that shared perception about blind spots is fundamentally unverifiable without trusted infrastructure. That would still be a useful result.
References
- Pretend Benign: A Stealthy Adversarial Attack by Exploiting Vulnerabilities in Cooperative Perception
- Occlusion-Aware Risk Assessment for Autonomous Driving in Urban Environments
- CP-Guard: Malicious Agent Detection and Defense in Collaborative Bird's Eye View Perception
- The Latent Guardian: Defending Collaborative Perception via Feature-Level Consistency Verification
- Research Challenges for the Security of Control Systems
- Secure Safety Filter: Towards Safe Flight Control under Sensor Attacks
- Combating False Reports for Secure Networked Control in Smart Grid via Trustiness Evaluation
- Safe Motion Planning in Unknown Environments: Optimality Benchmarks and Tractable Policies
- Behavior Planning of Autonomous Cars with Social Perception
- SparseCoop: Cooperative Perception with Kinematic-Grounded Queries
- CoDynTrust: Robust Asynchronous Collaborative Perception via Dynamic Feature Trust Modulus
- Reasoning Multi-Agent Behavioral Topology for Interactive Autonomous Driving
- OPV2V: An Open Benchmark Dataset and Fusion Pipeline for Perception with Vehicle-to-Vehicle Communication
- V2VNet: Vehicle-to-Vehicle Communication for Joint Perception and Prediction