PhysPatch: A Physically Realizable and Transferable Adversarial Patch Attack for Multimodal Large Language Models-based Autonomous Driving Systems
Authors: Qi Guo, Xiaojun Jia, Shanmin Pang, Simeng Qin, Lin Wang, Ju Jia, Yang Liu, Qing Guo
Organizations: School of Software Engineering, Xi’an Jiaotong University, Xi’an, China · Center for Frontier AI Research, A*STAR, Singapore · School of Computer Science and Engineering, Nanyang Technological University, Singapore · Northeastern University, China · Hangzhou Dianzi University, China · Southeast University, China
Multimodal Large Language Models (MLLMs) are becoming integral to autonomous driving (AD) systems due to their strong vision-language reasoning capabilities. However, MLLMs are vulnerable to adversarial attacks, particularly adversarial patch attacks, which can pose serious threats in real-world scenarios. Existing patch-based attack methods are primarily designed for object detection models and perform poorly when transferred to MLLM-based systems due to the latter's complex architectures and reasoning abilities. To address these limitations, we propose PhysPatch, a physically realizable and transferable adversarial patch framework tailored for MLLM-based AD systems. PhysPatch jointly optimizes patch location, shape, and content to enhance attack effectiveness and real-world applicability. It introduces a semantic-based mask initialization strategy for realistic placement, an SVD-based local alignment loss with patch-guided crop-resize to improve transferability, and a potential field-based mask refinement method. Extensive experiments across open-source, commercial, and reasoning-capable MLLMs demonstrate that PhysPatch significantly outperforms prior methods in steering MLLM-based AD systems toward target-aligned perception and planning outputs. Moreover, PhysPatch consistently places adversarial patches in physically feasible regions of AD scenes, ensuring strong real-world applicability and deployability.
Figures & tables
Figure 1: Overview of Differences Between PhysPatch and Existing Adversarial Patches: GAP ( Brown et al. 2017 ) , DAPatch ( Chen et al. 2022 ) , IAP ( Wei et al. 2024b )
Figure 2: Overview of the PhysPatch Framework: Semantic-based Mask Initialization, Global and Local Alignment Loss Calculation, and Mask and Patch Content Update
Methods
LLaVA-v1.6-13B
Qwen2.5-VL-72B
Llama-3.2-90B
GPT-4o
GPT-4.1
Claude-sonnet-4
ASR
AvgSim
ASR
AvgSim
ASR
AvgSim
ASR
AvgSim
ASR
AvgSim
ASR
AvgSim
Clean
0.0
0.103
0.0
0.092
0.0
0.101
0.0
0.101
0.0
0.099
0.0
0.102
IAP
5.0
0.164
0.3
0.120
2.0
0.146
1.5
0.126
0.1
0.108
0.3
0.114
DAPatch
8.8
0.180
1.5
0.129
2.4
0.150
1.9
0.134
0.3
0.118
0.6
0.197
AttackVLM
8.3
0.177
1.0
0.122
2.0
0.143
1.7
0.130
0.2
0.114
0.2
0.110
SSA-CWA
9.5
0.191
2.7
0.151
2.9
0.155
3.0
0.159
1.5
0.132
0.8
0.118
Table 1: Comparison of ASR and AvgSim Across Different Attacks on Various MLLMs. The best results are in bold.
Defense
Methods
LLama-3.2
GPT-4o
Claude-4
Gemini-2.0
GPT-o3
QVQ-Plus
ASR
AvgSim
ASR
AvgSim
ASR
AvgSim
ASR
AvgSim
ASR
AvgSim
ASR
AvgSim
Gaussian
FOA-Attack
31.9
0.342
32.1
0.359
9.5
0.160
21.4
0.287
13.6
0.188
18.9
0.259
PhysPatch
35.7
0.361
38.5
0.364
12.0
0.191
24.9
0.291
15.3
0.200
25.4
0.306
JPEG
FOA-Attack
28.2
0.329
27.4
0.312
9.2
0.154
21.8
0.280
12.9
0.171
17.2
0.243
PhysPatch
33.6
0.355
33.9
0.353
11.8
0.187
24.2
0.298
14.8
0.195
24.3
0.293
DISCO
FOA-Attack
27.6
0.320
28.0
0.327
8.9
0.148
19.2
0.271
12.0
0.163
15.9
0.235
Table 2: Robustness Comparison of PhysPatch and FOA-Attack Under Various Defense Mechanisms. Claude-4 refers to Claude Sonnet 4; this naming is used consistently throughout.
Methods
FID
LPIPS
BRISQUE
Time(s)
IAP
26.34
0.0224
50.89
148
DAPatch
7.52
0.0146
45.15
123
AttackVLM
4.85
0.0128
44.20
79
SSA-CWA
8.60
0.0125
45.22
1650
SIA
4.70
0.0111
46.20
806
MuMoDig
3.89
0.0108
45.20
924
Table 3: Comparison of Image Quality and Generation Time Across Different Attack Methods
Figure 3: Visualization of Perception in MLLM-Based AD Systems: Example from GPT-4o screenshot
Figure 4: Hyperparameter Sensitivity Analysis: SVD Dimension k , Weighting Factor η , Potential Step lr , Potential field diffusion coefficient σ , Init Threshold τ0 and Threshold Growth Rate β
Figure 5: Visualization Examples of Real-world Case Study
Methods
LLama
GPT-4o
Claude-4
Gemini-2.5
w/o mask update
34.3
37.9
13.3
22.9
w/o SVD (naive)
35.2
38.8
14.0
24.9
w/o patch crop
34.6
37.5
13.6
22.5
FOA Attack
33.1
34.3
13.4
21.0
Ours
37.2
40.3
14.5
25.4
Table 4: Ablation Results of Key Components in PhysPatch
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Rationale for excluding fine-tuned models such as DriveLM and Dolphin. These models are typically overfitted to specific datasets or scenarios, and therefore lack generalizability for evaluating real-world robustness.
Figure 7: Qualitative comparison of image quality across different adversarial patch attack methods. The patch regions are zoomed in for better observation.
Figure 8: Visualization results of attacking MLLM-Based autonomous driving systems. Taking perception attack as an example. We directly present the screenshots for display.
Model Type
Models
IAP
DAPatch
AttackVLM
SSA-CWA
SIA
MuMoDig
M-Attack
FOA-Attack
PhysPatch
Open Source
LLaVA-13B
4.6
7.4
6.3
7.9
8.4
9.0
25.2
29.1
38.5
Qwen-72B
0.1
1.2
1.0
1.9
2.6
2.4
13.5
13.7
15.0
LLama-90B
1.9
2.3
1.8
2.5
3.3
3.5
27.4
28.9
33.4
Commerical
GPT-4o
1.5
1.7
1.2
2.7
4.9
5.5
26.2
32.7
37.4
GPT-4.1
0.1
0.2
0.1
1.3
2.9
3.2
21.6
23.8
25.4
Claude-4
0.2
0.5
0.0
0.4
1.5
1.3
9.8
10.3
11.8
Appendix
Table 5: Comparison of Different Attack Methods on Planning Tasks across Various MLLMs. The best results are in bold.
Target
Methods
LLama-3.2
GPT-4o
Claude-4
Gemini-2.0
GPT-o3
QVQ-Plus
ASR
AvgSim
ASR
AvgSim
ASR
AvgSim
ASR
AvgSim
ASR
AvgSim
ASR
AvgSim
Speed Limit
FOA-Attack
36.0
0.349
30.0
0.332
14.0
0.200
26.0
0.316
17.0
0.233
22.0
0.284
PhysPatch
40.0
0.380
35.0
0.346
18.0
0.217
31.0
0.344
20.0
0.250
30.0
0.316
Pedestrian Crossing
FOA-Attack
34.0
0.344
35.0
0.372
14.0
0.197
22.0
0.289
20.0
0.255
23.0
0.275
PhysPatch
36.0
0.357
42.0
0.428
16.0
0.213
25.0
0.303
25.0
0.291
28.0
0.292
Appendix
Table 6: Comparison of PhysPatch and FOA-Attack across multiple adversarial targets under the perception task. The results are evaluated on 100 images for each target, demonstrating the effectiveness and generalization of PhysPatch on diverse semantic objectives.
Figure 9: Visualization of the transition from digital to physical space. (a) We selected 10 scenes covering diverse locations, viewpoints, and lighting conditions; (b) Digital-to-physical attack example: we visualize the original scene image, the digital adversarial sample, the printed adversarial patch, the physical adversarial sample, and the corresponding attack result.
Methods
LLama-3.2-90B-Vision
GPT-4o
Claude-sonnet-4
Gemini-2.5-flash
ASR
AvgSim
ASR
AvgSim
ASR
AvgSim
ASR
AvgSim
FOA-Attack
35.67 ±1.952
0.369 ±0.0288
35.33 ±1.874
0.361 ±0.0263
11.67 ±1.310
0.194 ±0.0075
21.33 ±0.501
0.282 ±0.0139
PhysPatch
42.33 ±2.081
0.405 ±0.0306
39.67 ±1.855
0.393 ±0.0281
13.67 ±1.527
0.217 ±0.0089
24.33 ±0.577
0.303 ±0.0143
Appendix
Table 7: Mean and standard deviation of FOA-Attack and PhysPatch on the perception task with the “stop sign” as the adversarial target, evaluated over 100 randomly sampled images with 3 independent runs.
Methods
LLama
GPT-4o
Claude-4
Gemini-2.5
Digital Result
IAP
0/10
0/10
0/10
0/10
FOA-Attack
3/10
4/10
0/10
1/10
PhysPatch
5/10
5/10
2/10
3/10
Physical Result
IAP
0/10
0/10
0/10
0/10
Appendix
Table 8: Comparison of attack success counts across digital and physical settings. We evaluate IPA, FOA-Attack, and PhysPatch on 10 scenes under both digital and physical conditions.
Patch Size
LLama
GPT-4o
Claude-4
Gemini-2.5
FID
90×90
38.0
33.0
13.0
21.0
3.14
120×120
42.0
36.0
14.0
26.0
3.72
150×150
47.0
42.0
14.0
30.0
5.07
Appendix
Table 9: Impact of patch size on ASR (%) across different MLLMs and corresponding FID score.
Trans.
Parameters
Remark
H Shift
[0,W/10]
Lateral Viewpoint Shift
V Shift
[0, H/10]
Vertical Viewpoint Shift
Rotation
±20∘
Camera Simulation
Scale
[0.25,1.25]
Distance/Resize
Noise
±16/255
Noise
Brightness
±0.1
Illumination
Appendix
Table 10: Transformation distributions used in EoT during adversarial patch generation. These transformations simulate various real-world factors such as viewpoint shifts, camera parameters, illumination changes, frequency filtering, and occlusions.
ϵ
LLama
GPT-4o
Claude-4
Gemini-2.5
FID
4
20.0
18.0
4.0
10.0
2.19
8
36.0
31.0
11.0
22.0
3.05
16
42.0
36.0
14.0
26.0
3.72
32
44.0
39.0
16.0
29.0
4.96
64
48.0
41.0
17.0
31.0
5.63
Appendix
Table 11: Impact of perturbation budge ϵ on ASR (%) across different MLLMs and corresponding FID score.