Pith. sign in

REVIEW 3 major objections 3 minor 58 references

When Cars Have Stereotypes: Auditing Demographic Bias in Objects from Text-to-Image Models

T0 review · 3 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Generated objects carry hidden demographic stereotypes, a new audit shows.

desk verdict The uploaded full text is an unrelated smart-home paper, so the SODA bias audit is unassessable; the abstract is intriguing but the manuscript is missing. read the letter →

arxiv 2508.03483 v3 pith:AWR6B6KT submitted 2025-08-05 cs.CV cs.AI

Dasol Choi , Jihwan Lee , Minjae Lee , Minsuk Kahng This is my paper
classification cs.CVcs.AI
keywords text-to-imagemodelsdemographicbiasstereotypesobjectgenerationauditframeworkvisualattributediscoverypromptdebiasing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that text-to-image models encode demographic stereotypes not only in images of people but also in images of everyday objects such as cars and laptops. The authors introduce SODA, a measurement framework that attaches demographic labels to object visuals and quantifies how strongly object appearance shifts with demographic cues in the prompt. They report that neutral prompts yield objects most visually similar to middle-aged White people, and that adding a demographic cue makes all 20 generated images share the same attribute value in 26.6% of object–model–demographic combinations. The paper argues these patterns make implicit associations measurable and offers SODA as a practical auditing pipeline.

What carries the argument

The central mechanism is the three-metric SODA framework built on automated attribute discovery. Rather than predefining which visual attributes matter, SODA first identifies attributes that actually vary across generated images, then uses them to compute BDS to measure divergence from the neutral baseline, CDS to measure disparity between demographic groups, and VAC to measure within-group diversity. This attribute-first design is what allows the audit to be applied across different object categories and models without manual bias annotation.

What would settle it

Recompute the BDS and CDS metrics using a different visual embedding model or human annotators to score demographic similarity, and check whether the reported pattern of middle-aged and White defaults and 26.6% homogeneity persists; if the similarity scores are dominated by embedding artifacts, the specific bias findings would change.

Watch

Extended reading notes

Core claim

The paper's central claim is that demographic bias in text-to-image models extends to objects, where it can be measured automatically. SODA works by first discovering the visual attributes that actually vary in generated images, then scoring each image's similarity to reference images of people from different demographic groups. Three metrics capture the bias: Base vs. Demographic Divergence (BDS) measures how far a demographic cue moves object appearance from the neutral default; Cross-Demographic Disparity (CDS) measures how differently two demographic cues style the same object; and Visual Attribute Concentration (VAC) measures how homogeneous the attribute values are within a set of images. Applying SODA to 8,000 images from five state-of-the-art models and eight object categories, the paper finds that neutral prompts already skew toward middle-aged and White defaults, that demographic cues trigger highly skewed stereotypical outputs with complete homogeneity in 26.6% of combinations, and that prompt-level debiasing reduces inter-group disparity but collapses within-group diversity, effectively replacing one stereotype with another.

Load-bearing premise

The audit assumes that an image of an object can be meaningfully assigned a demographic similarity score, that is, that visual features of a car or laptop can validly be mapped to human age and race categories.

Editorial extensions

If this is right

  • If SODA's measurements are correct, model developers gain a concrete checklist for auditing object-level bias before deployment, not just human-depiction bias.
  • The finding that neutral prompts already produce middle-aged and White defaults implies that even users who write purely descriptive prompts about objects receive outputs shaped by implicit demographic associations.
  • The 26.6% complete-homogeneity result implies that demographic cues can erase within-group diversity entirely, so a single attribute value (e.g., rose gold for laptops aimed at women) saturates the entire output set.
  • The finding that prompt-level debiasing reduces disparity but collapses diversity suggests that current debiasing interventions trade one form of bias for another, pointing to the need for interventions at the training or embedding level.
  • The framework could be extended to audit other axes of bias, such as age, disability, or cultural context, wherever objects can be scored for similarity to human demographic reference groups.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to test whether the demographic default skew toward middle-aged White people reflects training-data frequency distributions of depicted objects paired with people, which would make object bias a downstream symptom of human-depiction bias rather than an independent phenomenon.
  • The reliance on an embedding model to score demographic similarity means the audit measures the embedding's notion of demographic resemblance; a future comparison across different embedding models could determine how much of the reported bias pattern is stable and how much is an artifact of a particular visual-semantic space.
  • One testable prediction is that object categories more strongly associated with gendered or racialized human contexts (e.g., beauty products vs. power tools) will show higher CDS scores, a pattern that would connect the framework to broader stereotype literature.
  • The framework's attribute-first design suggests it could be repurposed as a general-purpose diversity audit for any generative domain, measuring not demographic stereotypes but any target attribute whose distribution should be broad, such as style, color, or form factor.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The manuscript is submitted under the title "When Cars Have Stereotypes: Auditing Demographic Bias in Objects from Text-to-Image Models" and its abstract describes SODA, a framework for measuring demographic bias in generated objects via the BDS, CDS, and VAC metrics. The abstract reports three headline empirical findings: neutral prompts produce outputs most visually similar to middle-aged and White people, 26.6% of object-model-demographic combinations collapse to a single attribute value across all 20 generated images, and prompt-level debiasing reduces inter-group disparity while shrinking within-group diversity. However, the full text supplied is an entirely different paper, "Semantic-aware Graph-guided Behavior Sequences Generation with Large Language Models for Smart Homes" (arXiv:2508.03484), which concerns LLM-based synthesis of smart-home behavior sequences and contains no SODA content. Because the body of the submission does not describe the SODA methodology, dataset, models, or metrics, the paper's central empirical claims cannot be assessed from the submitted artifact.

Significance. If the SODA findings were supported by a proper methodology, the paper would address a genuinely underexplored problem: text-to-image bias audits have concentrated on human depictions, while object-level demographic associations remain less studied. The reported 26.6% collapse rate and the claim of implicit middle-aged and White defaults are falsifiable and would be practically relevant for responsible AI evaluation. The three proposed metrics, if rigorously defined and validated, could be a useful addition to the audit toolbox. However, none of these contributions can be evaluated on the merits here: the submitted artifact contains no code, no data, no metric definitions, and no experimental protocol for SODA, so there is no reproducible or verifiable content to credit. The only assessable material is the unrelated SmartGen paper, which is not the work under review.

major comments (3)
  1. [Full text (Sections 1-6)] The body of the submission is the SmartGen paper on LLM-based smart-home behavior sequence synthesis, not the SODA audit announced in the title and abstract. Sections 1 through 6 describe time-semantic splitting, sequence compression, graph-guided synthesis, and a two-stage outlier filter, with experiments on anomaly detection and behavior prediction; they contain no mention of SODA, text-to-image models, demographic bias, BDS, CDS, VAC, or any of the eight object categories named in the abstract. The entire evidence base for the paper's central claims is therefore absent from the submitted artifact, and the manuscript is internally inconsistent as submitted.
  2. [Abstract] The headline quantitative findings—the middle-aged and White default result and the 26.6% all-20-identical collapse rate—are stated without any methodological detail in the available material. The abstract does not specify how an object image is mapped to human age and race similarity scores, which embedding space or classifier is used, what the demographic anchor set is, how attributes are discovered, what the neutral prompt set is, or how the debiasing prompt protocol is constructed. These are not presentation issues; they are the measurement procedures needed to reproduce the reported statistics, and their absence means the central claims are unverifiable from the submission.
  3. [Abstract] The three metrics BDS, CDS, and VAC are named but never defined in the submitted text. In particular, BDS requires a mapping from a generated object image to a position on human demographic-similarity dimensions; the abstract gives no evidence that such a mapping is valid. If the similarity signal is an artifact of the embedding model, the "most visually similar to middle-aged and White people" finding would collapse, and the document provides no validation or sensitivity analysis to rule out that possibility. This is a load-bearing methodological gap, not a minor omission.
minor comments (3)
  1. [Header/metadata] The running header of the full text cites arXiv:2508.03484v1, while the submission under review is arXiv:2508.03483; the metadata and the body reference different manuscripts.
  2. [Title/author block] The title and author block of the full text (SmartGen) do not match the title and abstract of the submission (SODA), so even the identity of the paper is inconsistent across the submitted artifact.
  3. [Full text (figures and tables)] None of the figures or tables in the full text relate to the SODA audit; for example, Figure 2 is a SmartGen architecture diagram and Tables 2-5 report smart-home anomaly detection and behavior prediction results, not demographic bias measurements.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the submitted full text contains no SODA methodology or metric equations, so there is no derivation chain to audit; the abstract's claims are unverifiable in this artifact, which is a completeness issue, not circularity.

full rationale

The circularity analysis looks for steps where a claimed prediction reduces, by the paper's own equations or by self-citation, to its inputs. In this submission, the abstract describes SODA, BDS, CDS, VAC, an 8,000-image audit, and findings about demographic defaults and stereotype concentration, but the full text is an unrelated paper on SmartGen, an LLM-based smart-home behavior sequence synthesis framework. The SmartGen paper contains no description of SODA, no definitions or equations for BDS, CDS, or VAC, no demographic-similarity scoring procedure, and no experimental setup for text-to-image object generation. Consequently, there is no formal derivation chain in the submitted artifact that could be shown to be circular. The absence of the SODA methodology means the central empirical claims cannot be checked, but that is a correctness and completeness problem rather than a circularity finding. No fitted parameter is renamed as a prediction, no result is justified solely by a self-citation, and no uniqueness theorem is imported from the authors' prior work. The reader's provisional circularity score of 0.0 is therefore consistent with the available evidence: within the material provided, there are no equations or argumentative steps that reduce to their own inputs.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The abstract describes an empirical audit but provides no visible free parameters or invented entities. The claims rest on domain assumptions about the validity of demographic similarity scoring for objects and the completeness of automatic attribute discovery.

assumptions (3)
  • domain assumption Demographic labels (age, race) can be meaningfully assigned to generated object images via visual similarity.
    The BDS metric and the finding that neutral prompts produce objects 'most visually similar to middle-aged and White people' require this mapping; the abstract does not justify it.
  • domain assumption Automated attribute discovery captures the attributes relevant to stereotypes without omitting or duplicating key visual dimensions.
    The VAC metric and the 26.6% collapse statistic depend on the discovered attribute taxonomy; no validation is visible in the abstract.
  • domain assumption The five chosen models and eight object categories form a representative sample for generalizing about text-to-image bias.
    The abstract generalizes across models and object categories based on this sample; the full methods were unavailable to assess selection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Cars Have Stereotypes: Auditing Demographic Bias in Objects from Text-to-Image Models." pith.science (2026). https://pith.science/paper/AWR6B6KT

@misc{pith2026250803483,
  author       = {Pith},
  title        = {Pith review of: When Cars Have Stereotypes: Auditing Demographic Bias in Objects from Text-to-Image Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AWR6B6KT}},
  note         = {Machine review of arXiv:2508.03483}
}
read the original abstract

While prior research on text-to-image generation has predominantly focused on biases in human depictions, demographic bias in generated objects remains relatively underexplored. We introduce SODA (Stereotyped Object Diagnostic Audit), a novel framework for systematically measuring these biases through automated attribute discovery and three standardized metrics: Base vs. Demographic Divergence (BDS), Cross-Demographic Disparity (CDS), and Visual Attribute Concentration (VAC). Applying SODA to 8,000 images across five state-of-the-art models and eight object categories (e.g., cars), we find that "neutral" prompts produce outputs most visually similar to middle-aged and White people, suggesting these groups are implicitly over-represented in model defaults. Furthermore, demographic cues trigger highly skewed stereotypical outputs: 26.6% of object-model-demographic combinations produce results where all 20 generated images share the exact same attribute value (e.g., rose gold laptops for women). Finally, prompt-level debiasing reduces inter-group disparity but paradoxically collapses within-group diversity, replacing one stereotype with another. SODA offers a practical pipeline for making these implicit associations measurable, serving as a step toward more responsible AI development.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 39 canonical work pages

  1. [1]

    Noureddine Amraoui and Belhassen Zouari. 2021. An ml behavior-based security control for smart home systems. InRisks and Security of Internet and Systems: 15th International Conference, CRiSIS 2020, Paris, France, November 4–6, 2020, Revised Selected Papers 15. Springer, 117–130

    work page 2021

  2. [2]

    Jason W Anderson, Ken E Kennedy, Linh B Ngo, Andre Luckow, and Amy W Apon. 2014. Synthetic data generation for the internet of things. In 2014 IEEE International Conference on Big Data (Big Data) . IEEE, 171–176

    work page 2014

  3. [3]

    Zhangir Azerbayev, Hailey Schoelkopf, Keiran Paster, Marco Dos Santos, Stephen McAleer, Albert Q Jiang, Jia Deng, Stella Biderman, and Sean Welleck

  4. [4]

    Yassir Bendou, Amine Ouasfi, Vincent Gripon, and Adnane Boukhayma. 2025. ProKeR: A kernel perspective on few-shot adaptation of large vision-language models. In Proceedings of the Computer Vision and Pattern Recognition Conference . 25092–25102

    work page 2025

  5. [5]

    Ada Chen, Yongjiang Wu, Junyuan Zhang, Jingyu Xiao, Shu Yang, Jen-tse Huang, Kun Wang, Wenxuan Wang, and Shuai Wang. 2025. A Survey on the Safety and Security Threats of Computer-Using Agents: JARVIS or Ultron? arXiv preprint arXiv:2505.10924 (2025)

    arXiv 2025

  6. [6]

    Zhuo Chen, Zhuotao Jin, Di Luo, Marin Soljačić, et al . 2025. Mutual Infor- mation Scaling Law for Long-Context Language Modeling. arXiv preprint arXiv:2503.04725 (2025)

    work page arXiv 2025

  7. [7]

    Zhaomin Chen, Chai Kiat Yeo, Bu Sung Lee, and Chiew Tong Lau. 2018. Autoencoder-based network anomaly detection. In 2018 Wireless telecommu- nications symposium (WTS) . IEEE, 1–5

    work page 2018

  8. [8]

    Adriel Cheng. 2019. PAC-GAN: Packet generation of network traffic using generative adversarial networks. In2019 IEEE 10th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON) . IEEE, 0728–0734

    work page 2019

Show all 58 references
  1. [9]

    Zhangyu Cheng, Chengming Zou, and Jianwei Dong. 2019. Outlier detection using isolation forest and local outlier factor. In Proceedings of the conference on research in adaptive and convergent systems . 161–168

    2019

  2. [10]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, and et al. 2025. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv:2501.12948 [cs.CL] https://arxiv.org/abs/2501.12948

    2025 arXiv

  3. [11]

    2021.{HAWatcher}:{Semantics- Aware} anomaly detection for appified smart homes

    Chenglong Fu, Qiang Zeng, and Xiaojiang Du. 2021.{HAWatcher}:{Semantics- Aware} anomaly detection for appified smart homes. In 30th USENIX Security Symposium (USENIX Security 21) . 4223–4240

    2021

  4. [12]

    Yi Gao, Kaijie Xiao, Fu Li, Weifeng Xu, Jiaming Huang, and Wei Dong. 2024. ChatIoT: Zero-code Generation of Trigger-action Based IoT Programs.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies8, 3 (2024), 1–29

    2024

  5. [13]

    Tianbo Gu, Zheng Fang, Allaukik Abhishek, Hao Fu, Pengfei Hu, and Prasant Mohapatra. 2020. IoTGaze: IoT security enforcement via wireless context analysis. In IEEE INFOCOM 2020-IEEE Conference on Computer Communications. IEEE, 884– 893

    2020

  6. [14]

    Yunpeng Huang, Jingwei Xu, Junyu Lai, Zixu Jiang, Taolue Chen, Zenan Li, Yuan Yao, Xiaoxing Ma, Lijuan Yang, Hao Chen, Shupeng Li, and Penghao Zhao. 2024. Advancing Transformer Architecture in Long-Context Large Language Models: A Comprehensive Survey. arXiv:2311.12351 [cs.CL]...

    2024 arXiv

  7. [15]

    Hyunsik Jeon, Jongjin Kim, Hoyoung Yoon, Jaeri Lee, and U Kang. 2022. Ac- curate Action Recommendation for Smart Home via Two-Level Encoders and Commonsense Knowledge. In Proceedings of the 31st ACM International Con- ference on Information & Knowledge Management (Atlanta, GA,...

    2022

  8. [16]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In 2018 IEEE international conference on data mining (ICDM) . IEEE, 197–206

    2018

  9. [17]

    Haijiang Li, Cangqi Zhou, Jing Zhang, and Dianming Hu. 2025. Open-World Knowledge Augmentation for Zero-Shot Information Extraction in LLMs. In International Conference on Intelligent Computing . Springer, 304–315

    2025

  10. [18]

    Ruoyu Li, Qing Li, Qingsong Zou, Dan Zhao, Xiangyi Zeng, Yucheng Huang, Yong Jiang, Feng Lyu, Gaston Ormazabal, Aman Singh, et al. 2024. IoTGemini: Modeling IoT Network Behaviors for Synthetic Traffic Generation. IEEE Transactions on Mobile Computing (2024)

    2024

  11. [19]

    Miao Lin, Vincent W Zheng, and Shili Xiang. 2018. Sequential context modeling for smart devices by Collaborative Hidden Markov Model. In2018 IEEE 4th World Forum on Internet of Things (WF-IoT) . IEEE, 771–777

    2018

  12. [20]

    Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. 2008. Isolation forest. In 2008 eighth ieee international conference on data mining . IEEE, 413–422

    2008

  13. [21]

    Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics 12 (2024), 157–173

    2024

  14. [22]

    Qiang Liu, Shu Wu, Diyi Wang, Zhaokang Li, and Liang Wang. 2016. Context- aware sequential recommendation. In 2016 IEEE 16th International Conference on Data Mining (ICDM). IEEE, 1053–1058

    2016

  15. [23]

    Knud Lasse Lueth. 2018. State of the IoT 2018: Number of IoT devices now at 7B – Market accelerating. https://iot-analytics.com/state-of-the-iot-update-q1-q2- 2018-number-of-iot-devices-now-7b/

    2018

  16. [24]

    Shalmoly Mondal, Prem Prakash Jayaraman, Pari Delir Haghighi, Alireza Has- sani, and Dimitrios Georgakopoulos. 2022. Situation-aware IoT data generation towards performance evaluation of IoT middleware platforms. Sensors 23, 1 (2022), 7

    2022

  17. [25]

    OpenAI. 2024. GPT-4 Technical Report. CoRR abs/2303.08774 (2024). https: //doi.org/10.48550/ARXIV.2303.08774 arXiv:2303.08774

    2024 doi arXiv

  18. [26]

    Neha Patki, Roy Wedge, and Kalyan Veeramachaneni. 2016. The synthetic data vault. In 2016 IEEE international conference on data science and advanced analytics (DSAA). IEEE, 399–410

    2016

  19. [27]

    Lakshmanan Rakkappan and Vaibhav Rajan. 2019. Context-aware sequential recommendations withstacked recurrent neural networks. In The world wide web conference. 3172–3178

    2019

  20. [28]

    Hasan Redžović, Aleksandra Smiljanić, and Milan Bjelica. 2017. IP traffic genera- tor based on hidden Markov models. parameters 1, 2 (2017), 1

    2017

  21. [29]

    Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme. 2010. Factor- izing personalized markov chains for next-basket recommendation. InProceedings of the 19th international conference on World wide web . 811–820

    2010

  22. [30]

    Phillip Rieger, Marco Chilese, Reham Mohamed, Markus Miettinen, Hossein Fereidooni, and Ahmad-Reza Sadeghi. 2023. ARGUS: Context-Based Detection of Stealthy IoT Infiltration Attacks. In Proceedings of the 32nd USENIX Conference on Security Symposium (Anaheim, CA, USA) (SEC ’23...

    2023

  23. [31]

    Phillip Rieger, Marco Chilese, Reham Mohamed, Markus Miettinen, Hossein Fereidooni, and Ahmad-Reza Sadeghi. 2023. ARGUS:Context-Based Detection of Stealthy IoT Infiltration Attacks. In 32nd USENIX Security Symposium (USENIX Security 23). 4301–4318. Conference’17, July 2017, Wa...

    2023

  24. [32]

    Markus Ring, Daniel Schlör, Dieter Landes, and Andreas Hotho. 2019. Flow-based network traffic generation using generative adversarial networks. Computers & Security 82 (2019), 156–172

    2019

  25. [33]

    Amit Kumar Sikder, Leonardo Babun, Hidayet Aksu, and A Selcuk Uluagac

  26. [34]

    Selcuk Uluagac

    Amit Kumar Sikder, Leonardo Babun, Hidayet Aksu, and A. Selcuk Uluagac

  27. [35]

    Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommenda- tion via convolutional sequence embedding. In Proceedings of the eleventh ACM international conference on web search and data mining . 565–573

    2018

  28. [36]

    Wenxin Tang, Jingyu Xiao, Wenxuan Jiang, Xi Xiao, Yuhang Wang, Xuxin Tang, Qing Li, Yuehe Ma, Junliang Liu, Shisong Tang, et al. 2025. SlideCoder: Layout- aware RAG-enhanced Hierarchical Slide Generation from Design. arXiv preprint arXiv:2506.07964 (2025)

    2025 arXiv

  29. [37]

    In Proceedings of the 35th Annual Computer Security Applications Conference (San Juan, Puerto Rico, USA) (ACSAC ’19)

    Aegis: A Context-Aware Security Framework for Smart Home Systems. In Proceedings of the 35th Annual Computer Security Applications Conference (San Juan, Puerto Rico, USA) (ACSAC ’19). Association for Computing Machinery, New York, NY, USA, 28–41. https://doi.org/10.1145/335978...

  30. [38]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guil- laume Lample. 2023. LLaMA: Open and Efficient Foundation ...

    2023 doi arXiv

  31. [39]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in Neural Information Processing Systems (NIPS) 30 (2017)

    2017

  32. [40]

    Niek Tax. 2018. Human activity prediction in smart home environments with LSTM neural networks. In 2018 14th international conference on intelligent envi- ronments (IE). IEEE, 40–47

    2018

  33. [41]

    Juan Wang, Shirong Hao, Ru Wen, Boxian Zhang, Liqiang Zhang, Hongxin Hu, and Rongxing Lu. 2020. IoT-praetor: Undesired behaviors detection for IoT devices. IEEE Internet of Things Journal 8, 2 (2020), 927–940

    2020

  34. [42]

    Jincheng Wang, Zhuohua Li, Mingshen Sun, Bin Yuan, and John CS Lui. 2023. Iot anomaly detection via device interaction graph. In 2023 53rd Annual IEEE/IFIP International Conference on Dependable Systems and Networks (DSN) . IEEE, 494– 507

    2023

  35. [43]

    Yuxuan Wan, Yi Dong, Jingyu Xiao, Yintong Huo, Wenxuan Wang, and Michael R Lyu. 2024. MRWeb: An Exploration of Generating Multi-Page Resource-Aware Web Code from UI Designs. arXiv preprint arXiv:2412.15310 (2024)

    2024 arXiv

  36. [44]

    Binrui Wu, Shisong Tang, Fan Li, Bing Han, Chang Meng, Jingyu Xiao, and Jiechao Gao. 2025. Aligning and Balancing ID and Multimodal Representations for Recommendation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (Toronto ON, Cana...

    2025

  37. [45]

    Jingyu Xiao, Yuxuan Wan, Yintong Huo, Zhiyao Xu, and Michael R Lyu. 2024. Interaction2Code: How Far Are We From Automatic Interactive Webpage Gener- ation? arXiv preprint arXiv:2411.03292 (2024)

    2024

  38. [46]

    Pan Wang, Shuhang Li, Feng Ye, Zixuan Wang, and Moxuan Zhang. 2020. Pack- etCGAN: Exploratory study of class imbalance for encrypted traffic classification using CGAN. In ICC 2020-2020 IEEE International Conference on Communications (ICC). IEEE, 1–7

    2020

  39. [47]

    Jingyu Xiao, Zhiyao Xu, Qingsong Zou, Qing Li, Dan Zhao, Dong Fang, Ruoyu Li, Wenxin Tang, Kang Li, Xudong Zuo, et al. 2024. Make your home safe: Time-aware unsupervised user behavior anomaly detection in smart homes via loss-guided mask. In Proceedings of the 30th ACM SIGKDD ...

    2024

  40. [48]

    Jingyu Xiao, Qingsong Zou, Qing Li, Dan Zhao, Kang Li, Wenxin Tang, Runjie Zhou, and Yong Jiang. 2023. User Device Interaction Prediction via Relational Gated Graph Attention Network and Intent-aware Encoder. In Proceedings of the 2023 International Conference on Autonomous Ag...

    2023

  41. [49]

    Jingyu Xiao, Ming Wang, Man Ho Lam, Yuxuan Wan, Junliang Liu, Yintong Huo, and Michael R Lyu. 2025. Designbench: A comprehensive benchmark for mllm-based front-end code generation. arXiv preprint arXiv:2506.06251 (2025)

    2025

  42. [50]

    Yucheng Yin, Zinan Lin, Minhao Jin, Giulia Fanti, and Vyas Sekar. 2022. Practical GAN-based synthetic IP header trace generation using NetShare. In Proceedings of the ACM SIGCOMM 2022 Conference . 458–472

    2022

  43. [51]

    Zhouliang Yu, Ruotian Peng, Keyi Ding, Yizhe Li, Zhongyuan Peng, Minghao Liu, Yifan Zhang, Zheng Yuan, Huajian Xin, Wenhao Huang, et al. 2025. Formalmath: Benchmarking formal mathematical reasoning of large language models. arXiv preprint arXiv:2505.02735 (2025)

    2025 arXiv

  44. [52]

    Jingyu Xiao, Qingsong Zou, Qing Li, Dan Zhao, Kang Li, Zixuan Weng, Ruoyu Li, and Yong Jiang. 2023. I Know Your Intent: Graph-enhanced Intent-aware User Device Interaction Prediction via Contrastive Learning. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquito...

    2023

  45. [53]

    Dan Zhao, Qing Li, Qingsong Zou, Jingyu Xiao, Kaidong Wu, Ruoyu Li, Jianhui Lv, Yong Jiang, and Keqin Li. 2025. Security and privacy in smart homes: Challenges and latest developments. In Advances in the Internet of Things . CRC Press, 36–55

    2025

  46. [54]

    Qingsong Zou, Qing Li, Ruoyu Li, Yucheng Huang, Gareth Tyson, Jingyu Xiao, and Yong Jiang. 2023. Iotbeholder: A privacy snooping attack on user habitual behaviors from smart home wi-fi traffic. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies...

    2023

  47. [55]

    Dan Zhao, Qing Li, Qingsong Zou, Jingyu Xiao, Kaidong Wu, Ruoyu Li, Jianhui Lv, Yong Jiang, and Keqin Li. 2025. 2 Security and Privacy. Advances in the Internet of Things: Challenges, Solutions, and Emerging Technologies (2025), 36

    2025

  48. [58]

    Qingsong Zou, Jingyu Xiao, Qing Li, Zhi Yan, Yuhang Wang, Li Xu, Wenxuan Wang, Kuofeng Gao, Ruoyu Li, and Yong Jiang. 2025. QueryAttack: Jailbreaking Aligned Large Language Models Using Structured Non-natural Query Language. arXiv preprint arXiv:2502.09723 (2025). Semantic-awa...

    2025 arXiv

  49. [2019]

    In Proceedings of the 35th Annual Computer Security Applications Conference

    Aegis: A context-aware security framework for smart home systems. In Proceedings of the 35th Annual Computer Security Applications Conference . 28–41

  50. [2023]

    arXiv preprint arXiv:2310.10631 (2023)

    Llemma: An open language model for mathematics. arXiv preprint arXiv:2310.10631 (2023)

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.