PhysPatch: A Physically Realizable and Transferable Adversarial Patch Attack for Multimodal Large Language Models-based Autonomous Driving Systems

arXiv:2508.05167 · cs.CV · Submitted 2025-08-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "PhysPatch: A Physically Realizable and Transferable Adversarial Patch Attack for Multimodal Large Language Models-based Autonomous Driving Systems".

Jane: Here's the gist of the paper. 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 PhysPatch, a physically realizable and transferable adversarial patch framework tailored for MLLMbased AD systems, jointly optimizes patch location, shape, and content to enhance attack effectiveness and real-world applicability PhysPatch Framework Overview The proposed PhysPatch framework is designed to address the limitations of existing patch-based methods by jointly optimizing three key aspects of an adversarial patch: location, shape, and content To overcome weak attack effectiveness and poor transferability, the method replaces naive PGD-based optimization with a Global-Local Feature Alignment strategy This involves introducing a theoretically grounded SVD-based Local Alignment Loss inspired by optimal semantic compression principles Furthermore, it incorporates a Patch-Guided Crop-Resize Strategy to guarantee patch inclusion in every sampled crop, mitigating gradient vanishing issues Semantic Initialization and Refinement The method employs a Semantic-Based Mask Initialization strategy to identify physically deployable regions for patch placement in AD scenarios This initialization combines MLLM reasoning with potential field modeling Specifically, it utilizes SoM to extract semantic and spatial information, leverages GPT-4o to infer suitable patch placement regions Rj, and applies a region-centric potential field algorithm R to compute the centroid coordinate p and corresponding Gaussian potential field Φ The binary mask M is then generated by thresholding the potential field, with post-processing applied to enhance mask quality Transferability Enhancement Mechanisms To improve cross-model transferability, PhysPatch introduces a novel SVD-based local alignment loss and a patch-guided crop-resize strategy The SVD-based loss uses truncated SVD to achieve optimal semantic compression of local features and robustness to encoder variations This is achieved by forming representations f adv i and f tar i using left singular vectors U and singular values Σ from the rank-k SVD of local feature matrices The patch-guided crop-resize strategy, Tpatch, ensures the patch center p is included in every sampled crop by sampling the top-left corner (x, y) within constrained ranges based on a target crop area Ar Optimization and Real-World Applicability The optimization process involves an adaptive potential field update algorithm to refine the patch shape within physically feasible regions The mask is updated by computing the gradient G of the loss L with respect to M, updating Φ as Φ ← Φ + lr · max(0, G), and iteratively refining M based on P(Φ, τ) until a predefined threshold is met Extensive experiments demonstrate that PhysPatch consistently outperforms state-of-the-art (SOTA) methods in attack success rate and semantic alignment Moreover, the method reliably places adversarial patches in physically feasible regions of AD scenes, ensuring strong real-world applicability and deployability Experimental Validation Experiments were conducted across open-source, commercial, and reasoning-capable MLLMs on the nuScenes dataset The evaluation uses the LLM-as-a-Judge framework with GPT-4o to assess attack success rate (ASR) and average similarity (AvgSim) PhysPatch consistently outperforms existing SOTA methods across various model types, achieving superior results in both perception and planning tasks Furthermore, real-world case studies confirm the physical realizability of the attack, showing that PhysPatch successfully induces target-aligned responses in diverse driving environments Ablation and Trade-offs Ablation experiments confirm the necessity of each component; removing the potential-field-based mask update, SVD-based local alignment loss, or patch-guided cropping results in a noticeable decline in attack performance Hyperparameter studies show that while larger patches yield higher ASR, they lead to increased FID scores, illustrating a clear trade-off between attack effectiveness and image quality The method maintains high visual fidelity, with adversarial patches occupying no more than 1% of the entire image area PhysPatch proposes a physically realizable and transferable adversarial patch attack targeting MLLM-based AD systems by combining semantic-aware mask initialization, SVD-based local alignment, and patch-guided cropping The findings expose critical vulnerabilities in current MLLM-based AD systems and underscore the urgent need for robust physical-world defenses Future work plans to explore road-level scene understanding from a bird’s-eye view (BEV) perspective to expand potential placement areas The research emphasizes the need for responsible disclosure and collaborative research to mitigate potential risks associated with these techniques References Anthropic.

Jane: Introducing Claude 4.

Tom: https://www.anthropic.com/news/claude-4 Accessed 2025-06-16 Athalye, A.; Engstrom, L.; Ilyas, A.; and Kwok, K.

Title and authors: Tom: So the title is PhysPatch: A Physically Realizable and Transferable Adversarial Patch Attack for Multimodal Large Language Models-based Autonomous Driving Systems, which basically means they are tackling the challenge of making adversarial patches that actually work well on these massive vision language models for autonomous driving, but with a focus on real-world applicability.

Jane: It’s important that whatever they design can be physically realized and transferable across different models in this complex domain.

Lu: I find interesting how they are addressing the weakness of current patch attacks, which often don't transfer well from object detection models to these more complicated multimodal systems.

Meng: That’s a key hurdle for any research in this area; if an attack works on one model but fails on another, it doesn't help us deploy anything useful.

Lalam: I see how the authors are trying to create something that solves this by focusing specifically on MLLM-based autonomous driving systems rather than just general vision models.

The paper's summary: Tom: So the main point of PhysPatch is that they move away from just using standard patch optimization methods and instead use a global-local feature alignment strategy to improve how effective the attack is, which sounds like they are trying to make the attack smarter about where it puts the patch.

Jane: They introduce two big ideas for this: first, a SVD based local alignment loss to handle local feature redundancy, and second, they have this semantic-aware mask initialization combined with an adaptive potential field update algorithm for placing the patch in physically deployable spots.

Lu: It sounds like they are using the model's reasoning ability itself to suggest where the patch should go, combining things like using a model to get semantic information and then feeding that into a potential field algorithm that guides how the mask updates.

Meng: It’s about making sure the AI uses its understanding of what’s physically possible when deciding where to place the patch, which is much more grounded than random placement.

The paper's improvements: Tom: So the improvements they highlight are pretty focused on making it both better at fooling the model and better at being transferable across different models, specifically by using that SVD based local alignment loss which is inspired by optimal semantic compression principles to handle those local features.

Jane: And then they have this patch guided crop resize strategy, which guarantees the patch stays visible in every sampled crop, cutting out that gradient vanishing issue that happens with naive cropping methods.

Lu: That cropping thing sounds like a very practical fix for when the optimization process gets messy and you lose information during training.

Meng: It’s a good move if it solves those gradient vanishing issues because they are always a headache when you're trying to fine-tune things at scale.

Conclusion: Tom: So to wrap up, PhysPatch aims to be a physically realizable and transferable adversarial patch attack targeting MLLM based autonomous driving systems by combining semantic-aware mask initialization, SVD based local alignment, and patch guided cropping.

Jane: It shows that using the AI's reasoning capabilities to place the patch in feasible regions helps significantly with attack success rate across different models.

Lu: And it’s really about moving from just a theoretical paper to something that has real-world applicability for autonomous vehicles because of this approach.

Meng: From my side, I worry about how much we rely on that semantic initialization; if the reasoning model gets confused, the patch placement might be totally wrong in a real scene.

Lalam: I’m pretty sure this is solid because the results show it consistently outperforms state of the art methods in attack success rate and semantic alignment when tested on datasets like nuScenes.

Jane: It’s not just better, it reliably places adversarial patches in physically feasible regions of driving scenes, which is crucial for deployment.

Lu: Yeah, that reliability is what sells it for real-world use because the paper addresses the transferability problem head-on by using those SVD and cropping strategies.

Tom: PhysPatch is a solid piece of work because it addresses the transferability problem head-on by using those SVD and cropping strategies.

Jane: It leaves a clear path forward for making these attacks more robust against model changes, which is important when you are shipping these systems out to the road.

School of Software Engineering, Xi’an Jiaotong University · School of Computer Science and Engineering, Nanyang Technological University School of Computer Science and Engineering, Northeastern University, Hangzhou Dianzi University School of Software Engineering at Xi’an Jiaotong University Center for Frontier AI Research at A*STAR

cs.CV

Submitted: 2025-08-07

Updated: 2026-11-12

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: Multimodal Large Language Models (MLLMs) are becoming integral to autonomous driving (AD) systems due to their strong vision-language reasoning capabilities However…

Key concepts

Physically Realizable Adversarial Patch Framework
This framework optimizes three aspects of an adversarial patch: its location, shape, and content. It uses a semantic-based mask initialization and a potential field update to ensure the patches are placed in physically feasible regions of driving scenes, making the attack practical for real-world deployment.
SVD-based Local Alignment Loss
This loss function uses truncated Singular Value Decomposition (SVD) to achieve optimal semantic compression of local visual features. It helps make the generated adversarial patches robust against variations in different vision encoders used by various MLLMs, enhancing the attack's transferability.
Patch-Guided Crop-Resize Strategy
This strategy ensures that the chosen patch center is included in every sampled crop during optimization. This prevents gradient vanishing issues and guarantees that the adversarial patch remains visible and relevant across different image crops, improving overall attack reliability.

Terminology

Summary

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 PhysPatch, a physically realizable and transferable adversarial patch framework tailored for MLLMbased AD systems, jointly optimizes patch location, shape, and content to enhance attack effectiveness and real-world applicability

PhysPatch Framework Overview

The proposed PhysPatch framework is designed to address the limitations of existing patch-based methods by jointly optimizing three key aspects of an adversarial patch: location, shape, and content To overcome weak attack effectiveness and poor transferability, the method replaces naive PGD-based optimization with a Global-Local Feature Alignment strategy This involves introducing a theoretically grounded SVD-based Local Alignment Loss inspired by optimal semantic compression principles Furthermore, it incorporates a Patch-Guided Crop-Resize Strategy to guarantee patch inclusion in every sampled crop, mitigating gradient vanishing issues

Semantic Initialization and Refinement

The method employs a Semantic-Based Mask Initialization strategy to identify physically deployable regions for patch placement in AD scenarios This initialization combines MLLM reasoning with potential field modeling Specifically, it utilizes SoM to extract semantic and spatial information, leverages GPT-4o to infer suitable patch placement regions Rj, and applies a region-centric potential field algorithm R to compute the centroid coordinate p and corresponding Gaussian potential field Φ The binary mask M is then generated by thresholding the potential field, with post-processing applied to enhance mask quality

Transferability Enhancement Mechanisms

To improve cross-model transferability, PhysPatch introduces a novel SVD-based local alignment loss and a patch-guided crop-resize strategy The SVD-based loss uses truncated SVD to achieve optimal semantic compression of local features and robustness to encoder variations This is achieved by forming representations f adv i and f tar i using left singular vectors U and singular values Σ from the rank-k SVD of local feature matrices The patch-guided crop-resize strategy, Tpatch, ensures the patch center p is included in every sampled crop by sampling the top-left corner (x, y) within constrained ranges based on a target crop area Ar

Optimization and Real-World Applicability

The optimization process involves an adaptive potential field update algorithm to refine the patch shape within physically feasible regions The mask is updated by computing the gradient G of the loss L with respect to M, updating Φ as Φ ← Φ + lr · max(0, G), and iteratively refining M based on P(Φ, τ) until a predefined threshold is met Extensive experiments demonstrate that PhysPatch consistently outperforms state-of-the-art (SOTA) methods in attack success rate and semantic alignment Moreover, the method reliably places adversarial patches in physically feasible regions of AD scenes, ensuring strong real-world applicability and deployability

Experimental Validation

Experiments were conducted across open-source, commercial, and reasoning-capable MLLMs on the nuScenes dataset The evaluation uses the LLM-as-a-Judge framework with GPT-4o to assess attack success rate (ASR) and average similarity (AvgSim) PhysPatch consistently outperforms existing SOTA methods across various model types, achieving superior results in both perception and planning tasks Furthermore, real-world case studies confirm the physical realizability of the attack, showing that PhysPatch successfully induces target-aligned responses in diverse driving environments

Ablation and Trade-offs

Ablation experiments confirm the necessity of each component; removing the potential-field-based mask update, SVD-based local alignment loss, or patch-guided cropping results in a noticeable decline in attack performance Hyperparameter studies show that while larger patches yield higher ASR, they lead to increased FID scores, illustrating a clear trade-off between attack effectiveness and image quality The method maintains high visual fidelity, with adversarial patches occupying no more than 1% of the entire image area

Conclusion and Future Work

PhysPatch proposes a physically realizable and transferable adversarial patch attack targeting MLLM-based AD systems by combining semantic-aware mask initialization, SVD-based local alignment, and patch-guided cropping The findings expose critical vulnerabilities in current MLLM-based AD systems and underscore the urgent need for robust physical-world defenses Future work plans to explore road-level scene understanding from a bird’s-eye view (BEV) perspective to expand potential placement areas The research emphasizes the need for responsible disclosure and collaborative research to mitigate potential risks associated with these techniques

References

Anthropic. 2025. Introducing Claude 4. p. 7 https://www. p. 7anthropic.com/news/claude-4 Accessed 2025-06-16

Athalye, A.; Engstrom, L. p. 7; Ilyas, A. p. 7; and Kwok, K. p. 7 2018. Synthesizing robust adversarial examples. p. 7 In International conference on machine learning, 284–293

Bai, S.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Song, S.; Dang, K.; Wang, P.; Wang, S.; Tang, J., et al. 2025. Qwen2. 5-vl technical report. p. 7 arXiv preprint arXiv:2502. p. 713923

Brown, T. B.; Mane, D.; Roy, A.; Abadi, M. p. 7; and Gilmer, J. p. 7, 2017. Adversarial patch. p. 0 arXiv preprint arXiv:1712. p. 709665

Caesar, H.; Bankiti, V. p. 7; Lang, A. H.; Vora, S.; Liong, V. p. 7 E.; Xu, Q.; Krishnan, A. p. 7; Pan, Y.; Baldan, G. p. 7, and Beijbom, O. p. 7, 2020. nuscenes: A multimodal dataset for autonomous driving. p. 7 In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11621–11631

Chen, Z.; Li, B.; Wu, S.; Xu, J.; Ding, S.; and Zhang, W. p. 0, 2022. Shape matters: deformable patch attack. p. 7 In European conference on computer vision, 529–548

Cheng, R.; Ding, Y.; Cao, S.; Duan, R.; Jia, X.; Yuan, S.; Wang, Z., and Jia, X. p. 0, 2024. PBI-Attack: Prior-Guided Bimodal Interactive Black-Box Jailbreak Attack for Toxicity Maximization. p. 7 arXiv preprint arXiv:2412. p. 705892

DeepMind, G., 2024. Introducing Gemini 2. p. 70: our new AI model for the agentic era. p. 7 Online

DeepMind, G., 2025. Gemini 2.5: Our most intelligent models are getting even better. p. 7 Online Multimodal LLM

Ding, G. W.; Wang, L.; and Jin, X. p. 5, 2019. AdverTorch v0. p. 7 1: An adversarial robustness toolbox based on pytorch. p. 7 arXiv preprint arXiv:1902. p. 707623

Dong, Y.; Chen, H.; Chen, J.; Fang, Z.; Yang, X.; Zhang, Y. p. 0 p. 0, et al., 2023. How robust is google’s bard to adversarial image attacks? p. 7 arXiv preprint arXiv:2309. p. 711751

Gao, S.; Jia, X.; Ren, X.; Tsang, I. p. 7; and Guo, Q. p. 0 p. 0, 2024. Boosting transferability in vision-language attacks via diversification along the intersection region of adversarial trajectory. p. 7 In European Conference on Computer Vision, 442–460

Gu, J.; Jiang, X. p. 7; Shi, Z.; Tan, H.; Zhai, X.; Xu, C.; Li, W., et al., 2024. A survey on llm-as-ajudge. p. 7 arXiv preprint arXiv:2411. p. 7 p. 715594

Guesmi, A.; Ding, R.; Hanif, M. p. 7 A.; Alouani, I. p. 7; and Shafique, M. p. 7, 2024. Dap: A dynamic adversarial patch for evading person detectors. p. 7 In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 24595–24604

Guo, Q.; Pang, S.; Jia, X.; Liu, Y.; and Guo, Q., 2024a. Efficient generation of targeted and transferable adversarial examples for vision-language models via diffusion models. p. 7 IEEE Transactions on Information Forensics and Security

Guo, X.; Zhang, R. p. 0; Duan, Y.; He, Y.; Nie, D.; Huang, W. p. 0, et al., 2024b. SURDS: Benchmarking Spatial Understanding and Reasoning in Driving Scenarios with Vision Language Models. p. 7 arXiv preprint arXiv:2411.13112

Han, D.; Jia, X.; Bai, Y.; Gu, J.; Liu, Y., and Cao, X. p. 7, 2023. Ot-attack: Enhancing adversarial transferability of visionlanguage models via optimal transport optimization. p. 7 arXiv preprint arXiv:2312. p. 704403

Heusel, M.; Ramsauer, H. p. 7; Unterthiner, T. p. 7; Nessler, B. p. 7; and Hochreiter, S. p. 7, 2017. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. p. 7 In Advances in Neural Information Processing Systems

Ho, C.-H.; and Vasconcelos, N. p. 5, 2022. Disco: Adversarial defense with local implicit functions. p. 7 Advances in neural information processing systems, 35: 23818–23837

Jia, X.; Gao, S.; Guo, Q.; Qin, S.; Ma., K.; Huang, Y. p. 0 p. 0, et al., 2025a. Semantic-Aligned Adversarial Evolution Triangle for High-Transferability VisionLanguage Attack. p. 7 IEEE Transactions on Pattern Analysis and Machine Intelligence

Jia, X.; Gao, S.; Qin, S.; Pang., T.; Du, C.; Huang, Y., et al., 2025b. Adversarial Attacks against Closed-Source MLLMs via Feature Optimal Alignment. p. 7 arXiv preprint arXiv:2505. p. 721494

Jing, L.; Wang, R.; Ren, W.; Dong, X.; and Zou, C. p. 7, 2024. PAD: Patch-agnostic defense against adversarial patch attacks. p. 7 In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 24472–24481

Li, J.; Li, D.; Xiong, C. p. 7, and Hoi, S. p. 7, 2022. Blip: Bootstrapping language-image pre-training for unified visionlanguage understanding and generation. p. 7 In International conference on machine learning, 12888–12900 PMLR

Li, Z.; Zhao, X.; Wu, D.-D.; Cui, J., and Shen, Z. p. 7, 2025. A frustratingly simple yet highly effective attack baseline: Over 90% success rate against the strong black-box models of gpt-4. p. 75/4o/o1. arXiv preprint arXiv:2503. p. 710635

Liu, H.; Li, C.; Li, Y.; Li, B.; Zhang, Y.; Shen, S., and Lee, Y. p. 7 p. 7 J., 2024. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge. p. 7 Accessed 2025-06-16

Liu, H.; Li, C.; Wu, Q., and Lee, Y. J., 2023. Visual instruction tuning. p. 7 Advances in neural information processing systems, 36: 34892–34916

Liu, J.; Levine, A. p. 7; Lau, C. P.; Chellappa, R. p. 7; and Feizi, S. p. 7, 2022. Segment and complete: Defending object detectors against adversarial patch attacks with robust patch detection. p. 7 In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 14973–14982

Ma, Y.; Cao, Y.; Sun, J.; Pavone, M. p. 8; and Xiao, C. p. 8, 2024. Dolphins: Multimodal language model for driving. p. 8 In European Conference on Computer Vision, 403–420

Madry, A.; Makelov, A. p. 8; Schmidt, L. p. 3; Tsipras, D. p. 8; and Vladu, A. p. 8, 2017. Towards deep learning models resistant to adversarial attacks. p. 8 arXiv preprint arXiv:1706. p. 806083

Mahendran, A.; and Vedaldi, A. p. 8, 2015. Understanding deep image representations by inverting them. p. 8 In Proceedings of the IEEE conference on computer vision and pattern recognition, 5188–5196

Mittal, A.; Moorthy, A. p. 4 K.; and Bovik, A. p. 4 C., 2012. NoReference Image Quality Assessment in the Spatial Domain. p. 8 IEEE Transactions on Image Processing, 21(12): 4695–4708

OpenAI, 2024. GPT-4o System Card. p. 8 https://openai. p. 8 p. 8com/index/gpt-4o-system-card/. p. 8 Accessed 2025-06-16

OpenAI, 2025a. Introducing GPT-4. p. 81 in the API. p. 8 https://openai.com/index/gpt-4-1/ Accessed 2025-06-16

OpenAI, 2025b. Introducing OpenAI o3 and o4-mini. p. 8 Online

Platforms, M., 2024. Llama 3.2: Revolutionizing edge AI and vision with open, customizable models. p. 8 https://ai. p. 7meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/. p. 8 Accessed 2025-06-16

Qwen, 2024. QVQ: To See the World with Wisdom. p. 8 https://qwenlm. p. 8github.io/blog/qvq-72b-preview/. p. 8 Accessed 2025-06-16

Radford, A.; Kim, J. W.; Hallacy, C. p. 8; Ramesh, A. p. 8; Goh, G.; Agarwal, S. p. 8; Sastry, G. p. 8, Askell, A. p. 8, Mishkin, P. p. 8, Clark, J. p. 8, et al., 2021. Learning transferable visual models from natural language supervision. p. 8 In International conference on machine learning, 8748–8763 PMLR

Ren, Y.; Zhao, Z.; Lin, C.; Yang, B.; Zhou, L.; Liu, Z., and Shen, C. p. 7, 2025. Improving integrated gradient-based transferable adversarial examples by refining the integration path. p. 8 In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 6731–6739

Schmidt, E., 1907. Zur Theorie der linearen und nichtlinearen Integralgleichungen. p. 8 Mathematische Annalen, 63(4): 433–476

Shin, R.; Song, D., et al., 2017. Jpeg-resistant adversarial images. p. 8 In NIPS 2017 workshop on machine learning and computer security, volume 1, 8

Sima, C.; Renz, K.; Chitta, K. p. 8; Chen, L.; Zhang, H. p. 0; Xie, C., Beißwenger, J. p. 8, Luo, P., Geiger, A. p. 8, and Li, H., 2024. Drivelm: Driving with graph visual question answering. p. 8 In European conference on computer vision, 256–274 Springer

Sreeram, S.; Wang, T.-H.; Maalouf, A. p. 8; Rosman, G. p. 8; Karaman, S. p. 8, and Rus, D. p. 8, 2024. Probing multimodal llms as world models for driving. p. 8 arXiv preprint arXiv:2405. p. 805956

Thys, S.; Van Ranst, W. p. 8; and Goedeme, T. p. 8, 2019. Fooling automated surveillance cameras: adversarial patches to attack person detection. p. 8 In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 0–0

Wang, L.; Zhang, T. p. 0; Qu, Y.; Liang, S. p. 8; Chen, Y.; Liu, A., et al., 2025. Black-box adversarial attack on vision language models for autonomous driving. p. 8

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Physically Realizable Adversarial Patch Framework: The PhysPatch framework jointly optimizes patch location, shape, and content to enhance attack effectiveness and real-world applicability by introducing a semantic-based mask initialization strategy for realistic placement and an adaptive potential field update algorithm to refine the patch shape within physically feasible regions.

  2. Enhanced Cross-Model Transferability: The system incorporates a SVD-based local alignment loss with patch guided crop resize to improve transferability, ensuring that adversarial examples generated on one MLLM can effectively mislead another MLLM in an autonomous driving system.

  3. Robust Semantic Alignment: The combination of global and local alignment losses ensures optimal semantic compression of visual features, making the attack more robust to architectural variations across different vision encoders.

  4. Real-World Deployability Assurance: By utilizing a patch-guided crop resize strategy that guarantees the inclusion of the patch in every sampled crop, PhysPatch mitigates gradient vanishing issues inherent in naive cropping methods and ensures adversarial patches are placed in physically feasible regions of AD scenes.

  5. Multi-Task Vulnerability Exploitation: The evaluation demonstrates that PhysPatch consistently outperforms state-of-the-art methods in steering MLLM-based AD systems toward target aligned perception and planning outputs, including successful attacks on complex tasks like planning (e.g., choosing the next action).

Sources

Related papers