Skip to content

Invisible to both cameras and LiDAR: the security of autonomous driving multi-sensor fusion perception under physical-world attacks

Invisible for both Camera and LiDAR: Security of Multi-Sensor Fusion based Perception in Autonomous Driving Under Physical-World Attacks

arXiv· arXiv:2106.09249· S&P 2021· Yulong Cao*, Ningfei Wang*, Chaowei Xiao*, Dawei Yang*, Jin Fang, Ruigang Yang, Qi Alfred Chen, Mingyan Liu, Bo LiUniversity of Michigan; University of California, Irvine; NVIDIA Research; Arizona State University; Baidu Research and National Engineering Laboratory of Deep Learning Technology and Application; Inceptio; University of Illinois at Urbana-ChampaignIndexed by this site · 21 days ago· Originally publishedSelected AttacksSelectedAttention41

Quick read

Attacks

Based on the paper PDF (AI-generated); refer to the original

The authors propose MSF-ADV, which uses a 3D-printable adversarial object to simultaneously attack the camera and LiDAR of autonomous driving multi-sensor fusion, achieving a success rate of 91%–100% on 4 MSF combinations and a 100% collision rate in simulation.

Threat modelWhite-box attack: the attacker fully understands the victim AD system's MSF algorithm and can collect camera images and LiDAR point clouds of the target road segment; the attack goal is to place a 3D-printed adversarial object so that the victim AV cannot detect the obstacle ahead and collides with it, requiring simultaneous attacks on both the camera and LiDAR perception channels.

Questions
Mass-produced autonomous driving generally adopts multi-sensor fusion (MSF) perception, whose design assumption is that the camera and LiDAR will not be attacked simultaneously. Prior research attacked only one of the two modalities separately, and how much safety assurance MSF provides as a defense strategy had never been tested.
Methods
The authors propose MSF-ADV: using shape-manipulable adversarial 3D objects as the attack carrier, through differentiable LiDAR/camera rendering and a soft point-inclusion approximation of cell-level aggregated features, jointly optimizing the object vertices with PGD to make both perception channels fail simultaneously.
Experiments and results
On 4 MSF combinations of Apollo and Autoware.AI and 100 KITTI scenes, the success rate is 91%–100%; the 3D-printed physical object evaded detection in 107/108 frames on a real-vehicle LiDAR; in LGSVL simulation, Apollo's collision rate against adversarial traffic cones is 100% (0% for benign).
Limitations and future work
The authors note that they did not conduct an end-to-end physical-world evaluation on real vehicles (only simulation and miniature-scale experiments), nor did they evaluate MSF algorithms based on DNN-based fusion; the defenses tested reduced the success rate to at most 66%, and are known to be bypassable by adaptive attacks.
Experimental setupA5-L、A2-L et al. · KITTI、LGSVL simulator et al. · attack success rate、LPIPS et al.
Evaluated models
A5-LA2-LA5-CYOLO v3 (Y3)
Benchmarks
KITTILGSVL simulatorBaidu ApolloAutoware.AI
Metrics
attack success rateLPIPS∆lp vertex distancevehicle collision ratebenign detection rate
AI summaryand selection rationaleResearchers from the University of Michigan, NVIDIA Research and other institutions propose in a paper the MSF-ADV attack, which uses a 3D-printable adversarial object to simultaneously deceive the camera and LiDAR perception of autonomous driving, making them unable to detect obstacles ahead; the paper states this is the first study targeting the security of multi-sensor fusion (MSF) perception. The attack starts from the 3D mesh of a normal object (such as a traffic cone) and uses differentiable rendering and soft point inclusion to approximately optimize the vertex shape; the finished product looks like a worn, damaged roadblock. Under the white-box setting where the attacker knows the algorithm, the method achieves a success rate exceeding 91% across 4 MSF combinations of Baidu Apollo and Autoware.AI, 3 object categories, and 100 real KITTI driving scenes each; the 3D-printed physical object escaped LiDAR detection in 99.1% of 108 frames; in production-grade simulation, the adversarial traffic cone caused the Apollo vehicle to have a 100% collision rate over 100 runs, versus 0% for a normal traffic cone.

Researchers from the University of Michigan, NVIDIA Research and other institutions propose in a paper the MSF-ADV attack, which uses a 3D-printable adversarial object to simultaneously deceive the camera and LiDAR perception of autonomous driving, making them unable to detect obstacles ahead; the paper states this is the first study targeting the security of multi-sensor fusion (MSF) perception. The attack starts from the 3D mesh of a normal object (such as a traffic cone) and uses differentiable rendering and soft point inclusion to approximately optimize the vertex shape; the finished product looks like a worn, damaged roadblock. Under the white-box setting where the attacker knows the algorithm, the method achieves a success rate exceeding 91% across 4 MSF combinations of Baidu Apollo and Autoware.AI, 3 object categories, and 100 real KITTI driving scenes each; the 3D-printed physical object escaped LiDAR detection in 99.1% of 108 frames; in production-grade simulation, the adversarial traffic cone caused the Apollo vehicle to have a 100% collision rate over 100 runs, versus 0% for a normal traffic cone.

Why it was selectedThe paper presents a printable adversarial object attack pipeline and success rates for simultaneously deceiving cameras and LiDAR, and the common defense assumption of multi-sensor fusion is challenged by the measurements.

Related events and news

1 events · 1 news mentions

News that mentions this paper

VicOne测试:被操纵输入可改变机器人行为

Paper analysis

6 questions; open one to read the full answer

  1. What problem does this paper attempt to solve?

    The authors study whether multi-sensor fusion (MSF) autonomous driving perception can be defeated by a single physical attack that simultaneously attacks the camera and LiDAR.

    Core question: whether the basic security assumption of multi-sensor fusion (MSF) autonomous driving perception—"not all perception sources will be attacked simultaneously"—holds, which had never been tested before.

    • Scenario and importance: mass-produced high-level autonomous driving (Waymo, Pony.ai, Baidu Apollo) commonly adopts camera+LiDAR MSF perception, and it is widely regarded as a general defense strategy against existing adversarial attacks (§I, §II-A). Perception failure directly affects safety decisions such as collision avoidance.
    • Gaps in existing research: prior physical-world attacks target only camera or LiDAR single-channel perception (stickers, posters, laser irradiation, etc.); the authors note that simply combining these designs cannot attack MSF (§III-B).
    • Core observation: shape changes of a 3D object can both change point positions in the LiDAR point cloud and change pixel values in the camera image, so shape manipulation can serve as a single attack vector that simultaneously attacks both channels (§IV-A).
    • Proposed in the paper: the attack method MSF-ADV, which models the attack as an optimization problem and generates 3D-printable adversarial objects (such as a traffic cone with a slightly worn appearance), causing the victim AV to fail to detect it and collide with it (§IV).
    • Threat model (§III-A):
      • Goal: cause MSF perception to miss the obstacle ahead and lead to a collision, requiring simultaneous attacks on both the camera and LiDAR channels.
      • Knowledge: white-box; the attacker fully understands the victim system's MSF algorithm (e.g., by reverse-engineering a purchased/rented victim vehicle model, or directly targeting an open-source system).
      • Capability: during the attack preparation phase, the attacker can collect camera images and LiDAR point clouds of the target road segment; at deployment, the 3D-printed object is placed on the road surface.
      • Victim system: MSF perception using camera+LiDAR fusion; the experiments target Baidu Apollo and Autoware.AI.
  2. What related research exists?

    Related work is divided into three groups: AD sensor and perception security, digital/physical-world adversarial attacks, and GN and GA as baselines.

    AD systems and perception security

    • Sensor-layer attacks—spoofing/jamming of cameras [40], [115], [10], [29] on LiDAR, and RADAR, ultrasonic, IMU attacks [40], [116]; the authors note these works are at the sensor layer, whereas this paper targets higher-level autonomous driving software (the "brain").
    • Perception-layer attacks—camera/LiDAR object detection [5], [6], [9], [10], [117], tracking [118], localization [119], lane detection [120]–[122], traffic light detection [123], end-to-end AD [12], [13]; the authors note they all attack only the camera or LiDAR single channel, and this paper is the first work to study MSF perception security (as claimed by the authors).
    • The assumption of MSF as a defense—[10], [27]–[29], [38]–[43], etc. treat MSF as a general defense strategy against existing attacks; the authors note that no one had previously tested whether this basic assumption can fail.

    Adversarial attacks

    • Digital-space attacks—[12], [50]–[56], [66], [85], [105], [124]–[127], etc.; the authors state this paper focuses on physical-world attack vectors.
    • Physical-world attacks—stickers/posters [5]–[9], adversarial patch [58], [59], EoT robust synthesis [60], etc.; the authors note they do not consider MSF, and per the analysis in §III-B, blindly combining these designs cannot directly achieve the goal due to three unique challenges.

    Baselines and comparison setup

    • Baseline methods: GN (shape perturbation adding Gaussian noise to vertices) and GA (genetic algorithm, using this paper's objective function as fitness but without differentiable approximation), used to test whether this paper's success comes from the optimization design or from the perturbation itself (§V-C).
    • Benchmark/data: the KITTI dataset with 100 real driving scenes; end-to-end evaluation uses the LGSVL simulator running Baidu Apollo (§V-A, §VI).

    The authors position this paper as: the first work to study MSF perception security and address its unique design challenges (differentiable rendering of both sensor channels, differentiable approximation of cell-level aggregated features) (§VIII, §IX, as claimed by the authors).

  3. How does the paper address this problem?

    MSF-ADV uses shape-controllable 3D objects as the attack vector, and performs end-to-end PGD optimization using differentiable rendering plus a soft point-inclusion approximation.

    MSF-ADV's overall approach: starting from the 3D mesh of a normal object, only vertex positions are moved to manipulate shape; differentiable rendering is used to simultaneously synthesize the attacked camera image and LiDAR point cloud, and a differentiable approximation bridges the computation of LiDAR's cell-level aggregated features, so that the entire attack can be optimized end-to-end with PGD (§IV-A, §IV-B).

    Problem formulation

    • The optimization objective is minSa Et∼ T[La(t(Sa); Rl, Rc, P, M) + λ · Lr(Sa, S)], where La is the adversarial loss that makes the MSF algorithm M miss the adversarial object S^a, Lr is the realizability loss that improves surface smoothness and printability, and T is the set of 3D transformations for EoT (Eq. (1)–(5), §IV-B).
    • The constraint Δ(Sa, S) ≤ ϵ limits the degree of shape deformation to ensure stealthiness, solved with L∞-distance PGD (Eq. (6)).

    Attack vector: adversarial 3D object

    • Shape manipulation is chosen over texture manipulation: shape affects both sensor modalities simultaneously; the authors deliberately do not change texture, holding that changing texture would harm stealthiness and cause printing problems (§IV-A).
    • The object imitates a traffic object that legitimately appears on roads and looks slightly worn (such as a traffic cone), and can be filled with heavy material to cause collision damage (§IV-A).

    Differentiable rendering (§IV-C)

    • Differentiable ray-casting [71] is used on the LiDAR side and NMR [72] on the camera side; S^a is defined in the LiDAR coordinate system and transformed to the camera coordinate system using the calibration matrix C to ensure physical consistency.

    Preprocessing approximation (§IV-D)

    • Key observation: all commonly used cell-level aggregated features in Table I (occupancy, count, height, intensity, density) can be derived in closed form from the point-inclusion property, which is inherently discontinuous and non-differentiable.
    • Proposes soft point-inclusion: trilinear interpolation is used to estimate the inclusion probability of a point in its nearest 8 cells (Eq. (7)), and tanh approximates the sign function to narrow the gap with the true value (Eq. (8)); Fig. 6 shows that the count feature under the tanh approximation is almost indistinguishable from the true value (per the authors).
    • All features can thus be derived differentiably, e.g. CNT(cm) = ∑PCi ∈ PC softPI(PCi, cm); occupancy uses sign(CNT) and approximates sign(x)=x in backpropagation.

    Objective function and robustness design (§IV-E)

    • Adversarial loss: for DNN-based fusion, directly minimize the MSF output confidence; rule-based fusion is non-differentiable, so instead minimize the sum of the confidences of the LiDAR and camera network outputs separately, so that the object does not appear in the fusion output regardless of the fusion rule.
    • Realizability loss: a Laplacian loss improves surface smoothness, plus a term preventing the object from being generated below the road surface (formulas in Appendix A).
    • Stealthiness: L∞ PGD constrains the per-vertex movement distance to less than ε.
    • Robustness: EoT applies random yaw rotation (±5°) and ground-position translation to S^a, to cover distance and angle changes as the victim vehicle approaches.
    • Key parameters (Table VIII): µ=100, λ=20, learning rate 0.001, PGD constraint 0.02, X sampling range (5, 35) m.
  4. What experiments did the paper conduct?

    On 4 MSF combinations and 100 KITTI scenes, the success rate is 91%–100%; physical and simulation validation show it can be deployed in reality and cause 100% collisions.

    Experimental setup

    • MSF combinations tested (§V-A): LiDAR models A5-L (Apollo v5.5), A2-L (Apollo v2.5) and camera models A5-C (Apollo v5.5), Y3 (YOLO v3, built into Autoware.AI) form 4 combinations; both systems use rule-based fusion.
    • Object types: traffic cone (0.5×0.5×1.0 m), bench (0.6×0.5×1.5 m), toy car (0.6×0.7×1.6 m), assigned according to the types supported by each combination; the toy car is used for all 4 combinations.
    • Scenes: 100 real driving scenes selected from KITTI per object type (one frame per scene), in which benign objects are 100% detectable; objects are placed 7 m in front of the victim vehicle by default (the braking distance at 25 mph).
    • Success criterion: success if and only if the adversarial object is simultaneously undetected by both the LiDAR and camera models of that combination; the authors note this is a lower bound on the success rate for the specific rule-based fusion logic.
    • Stealthiness metrics: average per-vertex perturbation distance ∆lp and LPIPS; user study with 105 Mechanical Turk participants.
    • Repetitions: 100 scenes per condition in the main experiments; 20 randomly sampled positions per position interval in the robustness experiments; 100 runs in simulation.

    Main results

    According to Table II (100 driving scenes, ε=2 cm):

    Swipe horizontally to read the table

    MSF combinationObjectSuccess rate∆l2 (cm)LPIPS
    A5-L + A5-CTraffic cone100%3.280.06
    A5-L + A5-CToy car91%3.460.02
    A5-L + Y3Bench100%3.390.20
    A5-L + Y3Toy car93%3.370.04
    A2-L + A5-CTraffic cone100%3.430.07
    A2-L + A5-CToy car96%3.340.17
    A2-L + Y3Bench100%3.300.20
    A2-L + Y3Toy car97%3.250.06
    • The authors state that MSF-ADV achieves a success rate of at least 91% across all object types and combinations, with traffic cones and benches at 100%.
    • The authors note that A5-L + A5-C has the lowest success rate, and consider the latest Apollo models the most robust among the 4 combinations.
    • Stealthiness: an average ∆l2 of about 3.4 cm suffices to achieve at least 91% success rate; average LPIPS is 0.10, which the authors state is comparable to the level of GAN-based image inpainting methods.

    Comparison with baseline attack methods (§V-C)

    • Setup: A5-L + A5-C, traffic cone; GN adds Gaussian noise with σ=2.1 cm, GA uses a population of 50 and generates 2000 candidates per scene (this paper's method uses 1000 iterations).
    • Results (Table IV): GN success rate 8% (∆l2 3.35 cm), GA success rate 9% (∆l2 1.84 cm), this paper's method 100% (∆l2 3.28 cm).
    • Authors' interpretation: the high success rate comes from the optimization design rather than the perturbation itself; Fig. 7 shows that the fitness reached by GA after 2000 trials is reached by this paper's method in 133 iterations, which the authors describe as 15 times more efficient.

    Robustness, physical objects and simulation

    • Position robustness (Table V, A5-L+A5-C, traffic cone): across three X-distance intervals of 5–35 m, the average success rate is 95.5%–96.6% with EoT and 79.2%–80.3% without EoT; the authors state EoT improves it by 20.5% on average.
    • Real-vehicle LiDAR experiment (§V-E1): a 3D-printed 75 cm adversarial cube (generated for A2-L), collected with a real-vehicle Velodyne HDL-64E; the adversarial cube is detected in only 0.9% (1 frame) of 108 frames, while the benign cube is detected in 84.8% (84 frames) of 99 frames.
    • Miniature experiment (§V-E2, Table VI): a 1:6.67 scale traffic cone, iPhone 8 Plus camera + VLP-16 LiDAR, 20 random positions; on A5-L+A5-C the benign detection rate is 95% (19/20) and the attack success rate is 90% (18/20); the transfer attack success rate against A2-L+A5-C is 85% (17/20).
    • End-to-end simulation (§VI): LGSVL + Apollo v5.0, vehicle speed 30 km/h, 100 runs each; the adversarial traffic cone collision rate is 100%, benign is 0%; the authors state the logs show the object evaded both channels' detection in every frame before fusion.

    Defense evaluation (§VII-B1)

    • Input transformations (Fig. 12, A5-L+A5-C, traffic cone): bit-depth reduction, median smoothing, JPEG compression, autoencoder reformation, four types; median smoothing is the most effective, reducing the attack success rate to 66% without affecting the benign detection rate; the authors note these methods are known to be bypassable by adaptive attacks.
    • Data augmentation AUG (Table VII, A5-L+Y3, bench): the attack success rate drops from 100% to 69%, the benign detection rate remains 100%, but the test set mAP drops from 44% to 32%.
    • Adversarial training (Y3, 2-step PGD approximation): after 900 epochs the success rate still reaches 95%; the authors consider 2-step PGD insufficient to generate effective adversarial examples and plan systematic future study.

    Other ablations and analyses

    • Stealthiness parameter ε (Table III, A5-L+A5-C, traffic cone): at ε=1.0 the success rate is 93% (∆l2 1.51 cm), at ε=0.5 it is 76% (∆l2 0.54 cm).
    • Parameter sensitivity (Table IX, 20 scenes): most sensitive to µ (75% at µ=10, 60% at µ=1000, 100% at the default 100); drops to 65% at λ=200; drops to 45% with a PGD initial point of 0.1.
    • Printability (Table X): 100% of generated objects pass PreForm and watertightness checks; the Laplacian loss reduces the self-intersection rate from 88.73% to 14.43% and curvature from 1.68 to 0.69, without affecting the success rate; using QECD alone reduces the success rate to 90%.
    • Transferability (§V-B): the average transfer attack success rate of the toy car across the 4 combinations is about 75% (as reported by the authors).
    • User study (Appendix B, Fig. 13, n=105): the proportion of adversarial traffic cones judged "to be a legitimate traffic cone" is similar to benign cones (about 60%), and the proportion selected as "most abnormal in shape" is consistently the lowest.
  5. What are some points that can be further explored?

    The authors list limitations such as the lack of end-to-end evaluation on real vehicles and the failure to test DNN-based fusion, and plan to research more effective defenses.

    Limitations and future directions noted by the authors

    • No end-to-end physical-world evaluation on real vehicles (§VII-A): due to cost and safety considerations, no end-to-end attack evaluation was conducted on real AVs; a production-grade simulator was used as a substitute.
    • DNN-based fusion not evaluated (§VII-A): the experiments target rule-based fusion; the effectiveness against DNN-based MSF algorithms remains unclear; the authors state this is not a limitation of the method itself (the design is general to both fusion types) and leave it for future work.
    • Future directions for defense research (§VII-B1): explore the effectiveness of input transformation defenses under MSF-ADV adaptive attacks; draw on the successful exploration of median smoothing to explore other input transformations; design more efficient adversarial training; explore extending certified robustness to 3D space and physical-world attacks.
    • Fusing more perception sources (§VII-B2): such as multiple cameras/multiple LiDARs at different positions, introducing RADAR; the authors note this cannot fundamentally defeat the attack (RADAR point clouds may also be affected by shape manipulation, and its detection algorithm is still a DNN), leaving it for systematic exploration.

    Experimental coverage

    • The MSF combinations tested are 4, from 2 open-source full-stack AD systems (Baidu Apollo, Autoware.AI), all rule-based fusion (§V-A).
    • The main experiments use 100 KITTI scenes per object type; robustness uses 20 positions per interval; the miniature experiment uses 20 positions; simulation uses 100 runs per condition (§V, §VI).
    • Physical-world experiments include: real vehicle with LiDAR only (HDL-64E, a 75 cm cube targeting A2-L); camera+LiDAR both sides use a 1:6.67 miniature setup (§V-E).
    • Defense evaluation covers 4 input transformations, data augmentation (Y3 only), and 2-step PGD adversarial training (Y3 only); certified robustness was not evaluated, and the authors state that their existing design targets 2D digital-space perturbations (§VII-B1).
    • The paper does not report details on the time or hardware cost of attack generation (only Table VIII gives optimization parameters).
  6. Summarize the main content of the paper

    The authors claim the first MSF perception security study: a 3D-printed adversarial object simultaneously fools the camera and LiDAR, with a 100% collision rate in simulation.

    This is an autonomous driving perception security paper; the authors claim it is the first security study targeting multi-sensor fusion (MSF) perception, directly challenging the basic MSF design assumption that "not all perception sources will be attacked simultaneously."

    • Problem: mass-produced AD systems commonly use camera+LiDAR fusion perception and regard it as a general defense, but prior attacks target only single-channel sensors, and how much security MSF provides is unknown.
    • Method: MSF-ADV uses shape-manipulable 3D-printed objects (such as a traffic cone with a slightly worn appearance) as the attack vector; through differentiable LiDAR/camera rendering and a soft point-inclusion approximation of cell-level aggregated features (tanh approximating sign), it jointly optimizes the vertices with PGD under an L∞ constraint, causing both perception channels to miss detection simultaneously; EoT improves robustness across distances and angles.
    • Main results: on 4 MSF combinations of Apollo and Autoware.AI and 100 KITTI scenes, the success rate for three object types is 91%–100%, with an average vertex perturbation of about 3.4 cm (Table II); the baselines GN and GA achieve only 8%–9% at the same perturbation magnitude (Table IV).
    • Physical objects and end-to-end: the 3D-printed cube evades detection in 107/108 frames on a real-vehicle LiDAR; in the miniature experiment the traffic cone achieves a 90% success rate against A5-L+A5-C (18/20); in LGSVL simulation, Apollo's collision rate against adversarial traffic cones over 100 runs is 100%, and 0% for benign.
    • Defenses: among the defenses tested, median smoothing reduces the success rate to at most 66%, and data augmentation to 69% (mAP drops from 44% to 32%); the authors note these defenses are known to be bypassable by adaptive attacks.
    • Authors' conclusion: MSF's security guarantee as a defense strategy is limited, and the authors hope this work drives the community to develop effective defenses; limitations include no end-to-end evaluation on real vehicles and no testing of DNN-based fusion.

Related papers

Read originalS&P 2021