Adversarial tracking: attacks on image sensors for autonomous vehicle object tracking
Adversarial Tracking: Attacks on Image Sensors for Object Tracking in Autonomous Vehicles
The authors evaluated the impact of image-specific and image-agnostic attacks against YOLOv3/YOLOv4 detectors on four TBD trackers, and the authors state that perturbing only the detector can reduce the HOTA of the YOLOv3 tracker from 53.45–56.19% to 14.85–17.03%.
The attacker adds adversarial perturbations with an l∞ norm constraint (ϵ=8) to the camera input images, attacking only the object detector in the TBD system without accessing the internals of the tracking module; there are two types: image-specific (perturbation generated per frame) and image-agnostic (a single UAP applied to all frames); the victim systems are vision-only tracking systems combining YOLOv3/YOLOv4 with SORT, Deep SORT, Strong SORT, and OC-SORT in CARLA simulation.
- Autonomous-driving Tracking-By-Detection systems rely on neural-network detectors, and detectors are susceptible to adversarial perturbations; the authors argue that detection accuracy alone cannot fully characterize the impact of an attack on downstream trajectory estimation, and that it needs to be systematically evaluated at the tracking level.
- On a 7-map dataset generated by the CARLA simulator, l∞-constrained image-specific attacks (per-frame iterative) and image-agnostic attacks (a single UAP) were applied to YOLOv3/YOLOv4 detectors, paired with four trackers—SORT, Deep SORT, Strong SORT, and OC-SORT—and tracking quality was measured with MOTA, HOTA, and IDF1.
- Under the image-specific attack, YOLOv3 trackers' MOTA drops to as low as -772.60% and HOTA to 14.85–17.03%; YOLOv4 is more robust but HOTA still drops to 25.06–26.86%; the UAP can also lower MOTA, and the authors state that perturbing only the detector is enough to break TBD tracking.
- The authors note that in the examples the false positives are mostly of the bird and chair classes, that vehicle-class filtering can intercept some false positives, that the sources of the remaining error are not decomposed, and that perturbations targeting the vehicle class are left for future comparison; the experiments are conducted only in the CARLA simulation environment, with two detectors and ϵ=8, and the paper does not report evaluation in real physical scenes or against adaptive defenses.
- Evaluated models
- YOLOv3YOLOv4SORTDeep SORTStrong SORTOC-SORT
- Benchmarks
- CARLA simulator(7 个地图,作者开源的 CARLA 跟踪数据集)
- Metrics
- MOTAHOTAIDF1
The paper evaluates adversarial attacks against autonomous-driving object-tracking systems in the CARLA simulator, testing combinations of YOLOv3 and YOLOv4 detectors with SORT, Deep SORT, Strong SORT, and OC-SORT respectively, and uses MOTA, HOTA, and IDF1 to measure tracking quality. Image-specific attacks reduced the HOTA of YOLOv3 trackers from 53.45-56.19% to 14.85-17.03%, with MOTA reaching -772.60% in the worst case; YOLOv4 trackers were more robust, with HOTA still dropping to 25.06-26.86%. Image-agnostic attacks require only a single universal adversarial perturbation (UAP) to degrade tracking accuracy, without needing to generate a perturbation separately for each frame.
Paper analysis
What problem does this paper attempt to solve?
The authors aim to answer: to what extent does perturbing only the object detector in a TBD system break multi-object tracking for autonomous driving.
Core question: how adversarial perturbations against the detector affect the tracking quality of Tracking-By-Detection (TBD) systems, which detection accuracy alone cannot fully characterize.
- Scenario and importance: Camera-based multi-object tracking (MOT) in autonomous driving estimates the positions and identities of surrounding objects; TBD systems combine a neural detector with motion prediction and data association, and their tracking performance depends on detector accuracy, while detectors are vulnerable to adversarial attacks (introduction).
- Insufficient existing understanding: Detection errors do not necessarily translate directly into tracking errors, because the data association stage may reject inconsistent detections; but persistent detection errors can fragment trajectories (Fig. 1). The authors believe a systematic evaluation at the tracking level is needed.
- Core hypothesis: The authors want to verify whether traditional statistical tracking components (Kalman filtering, the Hungarian algorithm) can mitigate the impact of adversarial attacks (Sections III, IV).
- Paper contributions: The authors open-sourced a CARLA-based dataset generation framework; established a tracking benchmark combining SOT (written as SOT in the main text, citing [9] as SORT), Deep-SORT, Strong-SORT, and OC-SORT with YOLOv3, YOLOv4, and YOLOX; evaluated two types of attacks, image-specific and image-agnostic (introduction contribution list).
- Threat model:
- Goal: Make the detector produce a different set of detection hypotheses for perturbed images, thereby breaking tracking (Equation (7)).
- Knowledge: The attack targets the detector (YOLOv3/YOLOv4) and requires computing gradients with respect to the input image; detector-only attacks do not require access to the internal tracking model (Section II-C).
- Capability: Add perturbations with an l∞ norm constraint (ϵ=8) to the input image; image-specific generates per frame, image-agnostic uses a single UAP covering all frames.
- Victim systems: Vision-only TBD tracking systems in CARLA simulation (four trackers × two detectors).
What related research exists?
Related work falls into three categories: MOT methods, tracking evaluation metrics, and adversarial attacks; this paper's positioning is to attack only the detector and systematically evaluate at the tracking level.
Multi-object tracking methods
- Sun et al. (2020) and Krebs et al. (2017)—reviewed multi-pedestrian tracking based on the TBD framework and deep-learning object tracking methods, respectively; these are background reviews cited in this paper.
- SORT (2016), Deep SORT (2017), Strong SORT (2023), OC-SORT (2023)—the four real-time online trackers evaluated in this paper; SORT uses Kalman filtering for state prediction, Deep SORT adds CNN appearance features, and OC-SORT does not rely on additional neural networks.
Tracking evaluation metrics
- MOTA (Bernardin and Stiefelhagen, 2008)—the CLEAR MOT metric combining FP, FN, and IDSW.
- IDF1 (Ristani et al., 2016)—measures the consistency between predicted identities and ground-truth identities.
- HOTA (Luiten et al., 2020)—balances detection accuracy (DetA) and association accuracy (AssA).
Adversarial attacks on tracking systems
- Jia et al. (ICLR 2020)—the paper states this is the first adversarial attack against MOT models, proposed in 2019, attacking a TBD method with YOLOv3 as the detector.
- Wiyatno and Xu (CVPR 2019)—physical adversarial textures against GOTURN-style single-object tracking; Lin et al. (2021)—attacked FairMOT and pedestrian trackers' appearance and distance cues to switch tracklets.
- Liu et al. (ICASSP 2022)—attacked association methods, changing object label evidence to break association; the paper notes that, unlike such methods, detector-only attacks change only the image input and do not touch the tracking module.
Attacks and defenses on detectors
- The paper states that YOLOX and CenterNet are two detectors commonly used in autonomous driving, and both are vulnerable to adversarial attacks (citing Zhou et al. CVPR 2023's F&F attack, Pang et al. 2024).
- Erkuş and Purutçuoğlu (2021)—temporal anomaly detection can serve as a potential defense idea for identifying adversarial perturbations.
Baselines and benchmarks: This paper does not compare other attack methods as baselines; the evaluation benchmarks are the authors' self-built CARLA dataset (7 maps) and the three metrics MOTA, IDF1, and HOTA, with the no-attack (No Attack) setting as the control.
Positioning: The authors distinguish this paper from work that attacks association or attacks tracklets, emphasizing that this paper perturbs only the detector input and systematically characterizes the attack's impact at the tracking level (rather than only at the detection level).
How does the paper address this problem?
Two types of PGD-style perturbations were applied to YOLOv3/YOLOv4 on the CARLA dataset, then tracking was run with four TBD trackers and measured with MOTA/HOTA/IDF1.
Overall approach: keep the TBD tracking pipeline fixed and attack only the object detector within it—using gradient ascent to generate perturbations that increase detection confidence and shrink predicted boxes, then observing the drop in downstream tracking metrics. The method has two variants: image-specific attack and image-agnostic attack (UAP).
Problem Setting and Formalization
- Detection assumption: for an input image x, the detector outputs N candidate detections H = {h1, ..., hN}, where each hi = (bi, ci, pi) contains a bounding box, a confidence, and K-class class probabilities (Eq. (5)).
- Attack goal: generate a perturbation δ ∈ [−1,1]^{whc} such that the detection output of the perturbed image x′ = x + δ differs, i.e., H(x′) ≠ H(x) (Eq. (7)).
- Adversarial loss: Ladv(x) = ∑i=1N σ(ci)σ(pi) / ∑i=1N [σ(biw)σ(bih)]2 (Eq. (10)), meaning to increase the confidence of all candidate boxes and shrink the predicted box sizes, thereby increasing the number of detections; maximizing this loss produces false boxes, and minimizing it suppresses detections.
Image-Specific Attack
- Update rule: PGD-style iterations δ_{t+1} ← Proj_{∞,ϵ}(δ_t + α_t·sign(∇_x L_adv(x + δ_t))) (Eq. (8)), where Proj constrains the perturbation within an l∞ radius ϵ, and the learning rate decays per iteration.
- Hyperparameters: attack strength ϵ = 8, learning-rate decay k = 0.99; updated per frame, 1 iteration per frame, 300 consecutive frames per map across 7 maps (Section IV-A).
Image-Agnostic Attack (UAP)
- Update rule: compute the adversarial gradient for each image in a sample image set X_S and accumulate updates to obtain a single universal perturbation applicable to multiple frames (Eq. (9)); the paper states that a smaller learning rate is used to improve cross-image transferability.
- Hyperparameters: trained on Map 02 for 100 iterations (the authors state Map 02 generalizes best across all maps), learning rate α = 0.001, decay k = 0.98, ϵ = 8 (Section IV-B).
Tracking Pipeline and Evaluation Method
- TBD components: detector outputs go through Kalman-filter state prediction (constant-velocity model, with state containing box center, scale, aspect ratio and velocity, Eq. (1)), optional CNN appearance features (Deep SORT), and Mahalanobis distance + Hungarian algorithm for data association (Section II-A).
- Combinations tested: four trackers SORT, Deep SORT, Strong SORT, OC-SORT × two detectors YOLOv3, YOLOv4; the introduction's contribution list also mentions YOLOX, but the experiments in Section IV report only YOLOv3 and YOLOv4.
- Metrics: MOTA (penalizes FP, FN, IDSW, Eq. (2)), IDF1 (identity consistency, Eq. (3)), HOTA (geometric mean of DetA and AssA, Eq. (4)); the paper explains that MOTA emphasizes detection and identity errors, IDF1 focuses on identity consistency, and HOTA balances detection and association.
- Environment: CARLA simulator, 7 maps; the dataset generation framework and attack code have been open-sourced by the authors (introduction footnote).
What experiments did the paper conduct?
Two types of attacks × two detectors × four trackers; under image-specific attack, YOLOv3's MOTA drops to as low as -772.60%.
Experimental Setup
- Models tested: detectors YOLOv3, YOLOv4; trackers SORT, Deep SORT, Strong SORT, OC-SORT (8 pairwise combinations in total).
- Dataset: an autonomous-driving dataset generated by the authors with the CARLA simulator, 7 maps in total.
- Attack settings: image-specific attack ϵ=8, k=0.99, 1 step per frame, 300 frames per map; image-agnostic attack trains a UAP on Map 02, 100 iterations, α=0.001, k=0.98, ϵ=8.
- Metrics: MOTA, HOTA, IDF1, all tracking-level metrics; the control is No Attack.
- Review method: metrics are computed automatically by definition; the paper does not use human or LLM review.
- Number of repetitions: the paper does not report repeated experiments or variance.
Main Results
According to Fig. 3 (MOTA, image-specific attack) and Fig. 4 (MOTA, image-agnostic attack):
Swipe horizontally to read the table
Detector Tracker No Attack Image-Specific Image-Agnostic YOLOv3 OC-SORT 61.10% -482.66% 45.45% YOLOv3 Strong-SORT 61.43% -772.60% 47.18% YOLOv3 Deep-SORT 57.24% -692.26% 43.28% YOLOv3 SORT 62.63% -525.81% 46.85% YOLOv4 OC-SORT 65.36% 17.07% 14.26% YOLOv4 Strong-SORT 63.25% 0.94% 16.95% YOLOv4 Deep-SORT 65.27% -6.30% 10.98% YOLOv4 SORT 60.43% 19.19% 14.70% - The authors state that the image-specific attack sharply lowers MOTA by producing many small false-positive boxes, and YOLOv3's MOTA becomes negative due to the many FPs; YOLOv4 produces fewer but larger false-positive boxes and is relatively more robust (Fig. 3 caption).
- The authors state that the UAP is generally less effective than the image-specific attack but can still lower MOTA; the drop for YOLOv3 is smaller than under the image-specific attack, while several YOLOv4 combinations show larger MOTA drops under the UAP, and the authors state that the relative effect depends on the detector (Section IV-B).
- According to the readings in Fig. 4, the MOTA of the YOLOv4 trackers under the UAP (10.98–16.95%) is lower than their values under the image-specific attack except for Deep-SORT, which is an exception to the generalization that "the UAP is generally weaker."
HOTA and IDF1 (Appendix Fig. 5, Fig. 6)
- HOTA: YOLOv3 drops from 53.45–56.19% under No Attack to 14.85–17.03% under the image-specific attack and 40.58–43.45% under image-agnostic; YOLOv4 drops from 56.31–59.49% to 21.96–26.86% (image-specific) and 24.11–26.32% (image-agnostic) (Fig. 5).
- IDF1: YOLOv3 drops from 69.30–74.31% to 9.95–14.09% (image-specific) and 55.20–62.43% (image-agnostic); YOLOv4 drops from 73.37–75.78% to 30.20–34.46% (image-specific) and 32.95–34.78% (image-agnostic) (Fig. 6).
- The authors state that the drops in HOTA and IDF1 are evidence of degraded trajectory association and identity preservation (Sections IV-A, IV-B).
False-Positive Class Analysis
- What was measured: the classes of false-positive boxes after the attack were observed (Fig. 4(a)(b), referred to in the text as example figures).
- Results: in the YOLOv3 examples the false positives are predominantly labeled bird, and in the YOLOv4 examples predominantly labeled chair.
- Authors' interpretation: vehicle-class filtering can reject many of these false-positive boxes, but the MOTA loss shows that downstream tracking is not fully protected; the authors also note that these examples cannot determine how much of the remaining error comes from missed vehicles or from accepted erroneous trajectories.
Other Ablations and Analyses
- Iteration steps: the authors state that increasing the number of optimization iterations per frame can improve the image-specific attack's effect, because the initial frames have fewer false positives and tracking is largely unaffected; the paper does not give specific iteration counts or corresponding numbers (Section IV-A).
- Detector comparison: YOLOv3 produces many small false-positive boxes, YOLOv4 produces fewer but larger false-positive boxes, and on this basis the authors state that YOLOv4 is relatively more robust (Fig. 3 caption).
- UAP training map: Map 02 was chosen to train the UAP, and the authors state that it generalizes best across all maps; the paper does not give numbers comparing generalization across maps (Section IV-B).
- Tracker comparison (no attack): the authors state that under the same detector, OC-SORT achieves tracking accuracy comparable to Strong SORT without an additional neural-network feature extractor (Section V).
What are some points that can be further explored?
The authors note that the false-detection categories and the sources of remaining error are not decomposed, and propose vehicle-targeted perturbations as a further comparison; the experiments cover only the simulation environment.
Limitations and future directions noted by the authors
- Error sources not decomposed (Section IV-B): The authors note that the false-detection examples cannot determine how much of the remaining error comes from missed vehicles or from accepted erroneous trajectories.
- Future direction: vehicle-targeted perturbations (Section IV-B): The authors state that vehicle-targeted perturbations would be a useful further comparison.
- The paper has no separate Limitations section; the above two items come from the authors' explicitly stated shortcomings and future work in the results discussion.
Scope of experimental coverage
- Combinations tested: 2 detectors, YOLOv3 and YOLOv4; 4 trackers, SORT, Deep SORT, Strong SORT, and OC-SORT; the introduction's contribution list mentions YOLOX, but the Section IV experiments do not report its results.
- Environment: CARLA simulator only, 7 maps; the image-specific attack runs 300 consecutive frames per map, with 1 iteration step per frame.
- Attack parameters: ϵ=8; the UAP is trained on Map 02 for 100 iterations; the paper does not report sweep experiments over other ϵ values or iteration budgets.
- Defense evaluation: The paper mentions temporal anomaly detection only in related work as a potential defense idea, and does not implement or evaluate any defense in the experiments; the main text mentions that vehicle-class filtering can reject some false detections, but does not give tracking metrics after filtering.
- Items not reported: The paper does not report the number of repetitions, confidence intervals, or statistical tests; it does not report the number of queries or the time cost of the attacks.
Summarize the main content of the paper
In CARLA simulation, perturbing only the detector can break all four TBD trackers, with YOLOv3's MOTA dropping to as low as -772.60%.
An adversarial-attack evaluation paper targeting autonomous-driving visual tracking, focusing on the tracking-level performance of Tracking-By-Detection (TBD) systems when the detector is perturbed.
- Problem: the tracking performance of TBD systems depends on neural detectors, and detection errors do not necessarily translate directly into tracking errors; the authors argue that the impact of adversarial attacks needs to be systematically characterized at the tracking level.
- Method: on a self-built CARLA dataset (7 maps), l∞-constrained (ϵ=8) image-specific attacks (per-frame PGD-style iterations) and image-agnostic attacks (a single UAP trained for 100 iterations on Map 02) are applied to YOLOv3/YOLOv4, paired with four trackers SORT, Deep SORT, Strong SORT, OC-SORT, and evaluated with MOTA, HOTA, IDF1.
- Main results: under the image-specific attack, YOLOv3 trackers' MOTA drops from 57.24–62.63% to -482.66% to -772.60%, HOTA drops from 53.45–56.19% to 14.85–17.03%, and IDF1 is as low as 9.95% (Fig. 3, 5, 6).
- Detector differences: YOLOv4 produces fewer but larger false-positive boxes, and the authors state it is more robust, but HOTA still drops to 21.96–26.86% (Fig. 5).
- UAP: a single universal perturbation can also lower MOTA (e.g., YOLOv4+Deep-SORT from 65.27% to 10.98%); the authors state it is generally weaker than the image-specific attack, but some YOLOv4 combinations are exceptions (Fig. 4).
- Authors' conclusion: perturbing only the detector is enough to break TBD tracking systems; under the same detector, OC-SORT matches Strong SORT without an additional neural network, and on this basis the authors argue that improving traditional components such as the Kalman filter can raise tracking accuracy without introducing a new attack surface; future work can compare targeted perturbations against the vehicle class.