DecisionRobust: Evaluating the Robustness of JEV-Style Decision Models on Text and Vision
DecisionRobust: Benchmarking Robustness for JEV Style Decision Models Across Text and Vision
DecisionRobust evaluates JEV-style decision models on text and images; the authors state that resistance to perturbation does not necessarily imply resistance to injection, and label checking lowers ASR.
- Decision models map instructions and observations to choices, judgments, or scores for routing, filtering, and approval. Inputs may be attacker-controlled, and outputs that conform to the format can still select the wrong action. Clean accuracy does not indicate robustness to manipulation.
- DecisionRobust preserves Choice, Noul, and Score, and constructs text and image perturbations plus four types of prompt injection. Candidates undergo label-consistency checks by external models and are then filtered by cross-model transfer; text edits additionally undergo semantic review and human decisions.
- The authors state attacks can cause errors beyond the filtering models, but multi-agent optimized image perturbations transfer weakly to many targets; robustness to perturbation does not imply robustness to injection, and label checking lowers the measured ASR.
- The authors note that only single decisions are evaluated, image perturbation transfer is limited, and computational constraints limit attack and task coverage. Text evaluation has 30 configurations, vision 17; curated text perturbations 141, injections 374.
- Evaluated models
- Jev (jev-1.13.0)KEVAutoTrustOpenJEVBosunLiquid D1Imajev-4BVisual-Jev-4BJev-OmniOneJev-4BStandardOne-8BGLiNER2.5-Decide-340MStrands Decider-2BPerplexity Decider v1.1-27BStandardOne-3BNeoHorse-Jev-4BOpenJEV-27BKEV-27BOpenAI Decisions API (gpt-6-luna)OneJev-27B
- Benchmarks
- AEGIS2.0BFCLJev Agent Skill RouterWorkflowEvalsNevIRESCIjevnavScreenSpotMultimodal-Mind2Webjev-arm-labJev Drive LabSiimImajevBenchFinQAGeometry3KSST-2MMLUPromptShieldSMS Spam
- Metrics
- sample-level ASRcandidate-level ASRflip ASRtarget ASRlabel-consistent ASR
The researchers release the DecisionRobust benchmark for systematically evaluating the robustness of JEV-style decision models against manipulation under text and visual inputs. The benchmark covers real-world decision scenarios, as well as semantically preserving text perturbations, image perturbations, and prompt injection attacks in both modalities. Experiments are conducted on 30 model configurations for text tasks (including 24 held-out models) and 17 configurations for vision tasks (including 12 held-out models). Results show that attacks can expose vulnerabilities beyond the models used for construction and filtering, but optimized image perturbations, even with multiple surrogate models, rarely transfer to most targets; relative model robustness varies with task and attack type, and resistance to input perturbation does not imply resistance to prompt injection; label consistency checks can substantially reduce measured attack success rates. The benchmark is publicly available on the project page.
Paper analysis
What problem does this paper attempt to solve?
JEV-style decision models map instructions and observations to structured decisions; clean accuracy does not indicate robustness to manipulation.
Decision models map instructions and observations to choices, judgments, or scores; clean accuracy does not indicate whether they still decide according to the task label under adversarial manipulation.
- Scenario: The authors state such models are used for skill selection, filtering, retrieval reranking, workflow routing, and approval. Inputs can be user requests, retrieved content, tool outputs, or images, and may be attacker-controlled; outputs can conform to the schema yet select the wrong action.
- Existing work: The authors believe robustness across application domains, modalities, and attack strategies, as well as transfer among decision models, remains poorly understood. The introduction states JevAdvBench tests fixed, single-query text editing.
- Observation: The authors state that meaning-preserving edits can still change tool selection, and instructions in retrieved content can steer decisions toward attacker-chosen labels. The authors believe that looking only at whether predictions change overestimates attack effectiveness.
- Proposed content: DecisionRobust preserves the task's own Choice, Noul, and Score interfaces and correctness criteria, evaluates under text and image perturbations and prompt injection, and constructs the evaluation set using validity checks and cross-model filtering.
- Threat model:
- Goal: Attacks target clean-correct decisions. Perturbations seek arbitrary errors; injections aim at predefined wrong answers.
- Control: Text perturbations modify only the specified observation, preserving the instruction and protected fields. Sensor attacks modify allowed numeric and categorical fields within task constraints. Image perturbations modify pixels and preserve the text request. Injections append instructions to the observation or question instruction, or embed them in images. Instruction injection assumes access to the task prompt, a stronger permission than observation injection. Model parameters, output schema, and reference answers are fixed.
- Access: Adaptive text attacks use the construction model's probabilities or scores, without parameter or gradient access. Optimized image attacks use surrogate gradients. Fixed injection templates and common image corruptions require no model feedback.
- Validity: Text perturbations should preserve task-relevant meaning and the reference decision; numeric and visual perturbations should preserve that decision under their respective constraints. Injections need not preserve semantics but should not add task-relevant evidence. Overlaid text in images must leave visible evidence still supporting the reference answer.
What related research exists?
Related work includes decision model evaluation, adversarial perturbations, and prompt injection; the authors position this paper as a robustness benchmark across tasks, modalities, and transfer.
Decision Models and Evaluation
- BERT (Devlin et al., 2019): Uses task-specific output layers and fixed labels for classification. The JEV the authors compare against accepts instructions, options, or scoring criteria at inference time without training an additional classification head, and returns choices, truth probabilities, and scores, rather than response text.
- WorkflowEvals and Deußer et al. (2026): The former examines application workflows; the latter studies task accuracy and calibration. The paper describes both without additional criticism.
- JevAdvBench (Hu et al., 2026): Measures adversarial change relative to clean decisions and measures rerun variance on the same request. The introduction states it tests fixed single-query text edits.
Adversarial Perturbations and Robustness Benchmarks
- Text and image attacks: The authors cite Lan et al. (2025), Li et al. (2020), Abad Rocamora et al. (2024), and Dong et al. (2018), stating that adversarial text edits and image perturbations can mislead classifiers and language models.
- AdvGLUE and RobustBench (Wang et al., 2021; Croce et al., 2021): Evaluate text perturbations and image perturbations, respectively.
- PromptRobust (Zhu et al., 2024): Evaluates LLM robustness to adversarial prompt perturbations.
Prompt Injection
- Perez and Ribeiro (2022) and Greshake et al. (2023): The former studies prompt injection techniques; the latter studies indirect prompt injection in external content.
- AgentDojo (Debenedetti et al., 2024): Evaluates prompt injection attacks and defenses for LLM agents in dynamic environments.
- CheckList (Ribeiro et al., 2020): Supplements aggregate accuracy with behavioral tests. The introduction lists it as testing beyond aggregate accuracy.
Baselines and Data
- Baselines: Appendix A.1 uses Repeat@8 (rerunning the original request 8 times) and Paraphrase@4 on the same clean-correct samples. Paraphrasing is done by Qwen3.5-122B-A10B-FP8, without seeing reference answers or using Jev feedback.
- Benchmarks/datasets: Table 1 includes AEGIS2.0, BFCL, Jev Agent Skill Router, WorkflowEvals, NevIR, ESCI, ScreenSpot, Multimodal-Mind2Web, ImajevBench, FinQA, Geometry3K, SST-2, MMLU, etc. The experimental source group also uses PromptShield, SMS Spam, and Siim.
The authors frame the difference from JevAdvBench as: judging correctness under text and visual attacks by task labels, and evaluating transfer to models not involved in candidate filtering.
How does the paper address this problem?
DecisionRobust preserves native interfaces, constructs perturbation and injection candidates, and evaluates after validity checks and cross-model filtering.
The overall pipeline is to construct attack candidates, perform validity checks, then filter by cross-model transfer, and finally evaluate on models not involved in filtering. The framework is named DecisionRobust. Figure 1 summarizes the pipeline.
Tasks and Decision Interfaces
- Coverage: Tasks come from official examples, public GitHub projects, and community applications, and adapt existing datasets (Table 1). Categories include content filtering, tool and skill routing, workflow evaluation, retrieval relevance, GUI grounding, physical state and motion, market direction, visual decision-making, and SST-2, MMLU language diagnostics.
- Three interfaces: Choice selects one option; Noul judges a proposition; Score scores on a given scale. Multi-class Noul asks one binary question per class, normalizes true-class probabilities, and takes the highest score. Classification and QA tasks provide corresponding options or propositions.
- Input adaptation: Text can be natural language, serialized tool descriptions, DOM, workflow records, or sensor observations. Images include synthetic scenes, screenshots, rendered financial tables, and geometric figures. Screenshot grounding is changed to selecting among marked elements; FinQA tables are rendered as images.
Attack Construction
- Text perturbation: The construction model is Jev (jev-1.13.0), and candidates are fixed before validation. Character-level uses DeepWordBug, Charmer; word-level uses TextFooler, SememePSO; sentence-level uses SCPN, PromptAttack S2/S3, StressTest. Sensor attacks modify values and specified category observations within field boundaries, and synchronize related summaries. The authors state that due to computational constraints, Charmer and SememePSO were not used on some data. Appendix E.1 gives method details and references.
- Text injection: Each question fixes one wrong target; four templates are Direct, Self-check, Note, Repetition (Table 2 gives examples), tested at two positions: observation and question instruction. Classification targets are deterministically selected from wrong answers; Score targets take the scale endpoint farthest from the reference. Original text, schema, and reference are preserved. Multi-class Noul instruction payloads push all questions toward the same target class.
- Image perturbation: The common five types are Gaussian noise, defocus blur, brightness change, JPEG compression, and resolution reduction, each with five levels. The first four use ImageNet-C parameters; resolution reduction downsamples first, then restores size with bicubic interpolation, without changing text. Optimized attacks minimize the sum of log probabilities of clean predictions from Imajev-4B, Visual-Jev-4B, and Jev-Omni for each image–question pair; use momentum updates, random scaling and padding, 200 steps, ℓ∞ budget 8/255, pixel clipping to [0, 1].
- Image injection: Reuses the four templates and fixed wrong targets. In-image injection renders outlined white text at four positions and three font sizes, 12 settings per template, with the question unchanged. Instruction injection appends the same payload to the question, preserving the clean image. Appendix E.1 has examples and rendering details.
Validation and Filtering
- Which candidates are checked: Perturbations and in-image injections undergo label consistency checks; text content injection and instruction injection skip this step. Then transfer filtering is done, followed by semantic review of text edits, or readability review of common image perturbations.
- Label consistency: Two external LLMs independently answer the perturbed task without seeing the reference, and both must recover the reference according to the task scoring rules. Text uses GPT-5 mini and DeepSeek-V4-Flash; images use Claude Sonnet 5. In-image injection instead has GPT-5 and Claude Sonnet 5 recover the reference from readable visual evidence, ignoring injected text; key answer content must remain visible. The authors plan to separately release the full candidate set that passed only applicable label checks, without transfer filtering and subsequent review.
- Filtering and human review: Prefer candidates that fool more filtering models. Text uses six, vision uses five filtering models, counting only clean-correct predictions on the same target question and format, and balancing task coverage. Other evaluation models do not participate in filtering. GPT-5 compares original text with edits, without seeing reference answers or model predictions; changes to facts, entities, quantities, relations, conditions, or intent are rejected; recoverable typos are allowed. Each text perturbation is decided by a human. Sensor perturbations only undergo label checks. Common image perturbations are human-checked for readability of task-relevant text and details.
Success Determination
Samples fix clean inputs, target questions, and decision formats. Choice uses native decisions. Score is correct when the API probability-weighted mean differs from the reference by no more than 0.5. ASR only uses samples where the model is clean-correct and the attack response is valid; unsupported tasks and invalid pairs are excluded.
Sample-level and candidate-level are ASRsample=(∑i∈ Dmax1≤ j≤ Kisij)/(|D|) and ASRcand=(∑i∈ D∑j=1Kisij)/(∑i∈ DKi). The former counts success by any candidate, the latter by each variant; candidates do not include intermediate search queries. Label-consistent ASR requires successful candidates to pass label checks, with the denominator unchanged. The above ratios for injection are flip ASR; target ASR instead counts hits on the predefined target, with the same denominator and Score tolerance 0.5.
What experiments did the paper conduct?
The authors state attacks can exceed the filtering models, optimized image perturbations transfer weakly to many targets, and robustness to perturbation does not bring robustness to injection.
Experimental Setup
- Models under test: 30 in the curated text set. Filtering used Jev (jev-1.13.0), KEV, AutoTrust, OpenJEV, Bosun, Liquid D1. The 24 not involved in filtering are Vega-27B, Lux-9B, NeoHorse-Jev-4B, GEV-26B, Mica-4B, Imajev-4B, Imajev-9B, OneJev-4B, OneJev-9B, OneJev-27B, Visual-Jev-4B, Jev-Omni, StandardOne-3B, StandardOne-8B, Strands Decider-2B, KEV-27B, OpenAI Decisions API (gpt-6-luna), H2O-Lightning-4B, Liquid d1-3B, Perplexity Decider v1.1-27B, InnerJev-27B-Full, GLiNER2.5-Decide-340M, CLM-v0.1-8B, Laya Multilingual-322M. 17 visual; filtering used Imajev-4B, Visual-Jev-4B, Jev-Omni, OneJev-4B, StandardOne-8B. The other 12 are StandardOne-3B, NeoHorse-Jev-4B, Strands Decider-2B, OneJev-9B, Imajev-9B, OneJev-27B, GEV-26B, OpenJEV-27B, OpenAI Decisions, H2O-Lightning-4B, Liquid d1-3B, Perplexity Decider v1.1-27B.
- Data scale: Before filtering, text perturbations are 10,093 candidates, 1,332 samples, 16 source groups; injections are 9,716 candidates, 1,232 samples, 15 source groups, with PromptShield excluded. Curated text is 141 perturbations (120 samples) and 374 injections (265 samples). Common image perturbations are 160 image–question pairs for Choice and Noul, totaling 320; optimized images are 160 Choice pairs. Curated images are 200 perturbations (127 common, 73 optimized), 223 in-image injections, 120 instruction injections.
- Attacks and baselines: Text perturbations transfer from Jev; recipients do not optimize further. Image optimization surrogates are Imajev-4B, Visual-Jev-4B, Jev-Omni. Controls are Repeat@8 and Paraphrase@4 (Qwen3.5-122B-A10B-FP8). Table 3's Combined ASR does not include these two baselines.
- Metrics: Sample-level and candidate-level ASR; injections additionally report flip ASR and target ASR. Checked or label-consistent requires successful candidates to pass label checking, with the denominator still using Raw's denominator. Score correctness threshold and target tolerance are both 0.5. Appendix C additionally reports clean/adversarial accuracy and reference-answer probability.
- Validation models: Text label checking uses GPT-5 mini and DeepSeek-V4-Flash; images use Claude Sonnet 5. In-image injections use GPT-5 and Claude Sonnet 5. Text semantic review uses GPT-5, and each text perturbation is decided by a human. Common image perturbations additionally undergo human readability checks. ASR itself is computed from decision outputs relative to task labels, not scored by these models.
- Denominators and repetition: Table 3's N is the pool size, from Security incidents' Choice N=3 to FSD and Siim's 120; the ASR denominator is each model's clean-correct and valid-response samples. "—" indicates no eligible samples. The paper does not report repetition counts for the main evaluation.
Main Results
The table below summarizes the six text models before filtering. The perturbation column is Table 3's sample-level Combined ASR (%) across the full dataset and all formats; the injection column is Table 4's sample-level flip/target ASR (%) for Either location. Jev is the text perturbation construction model; its perturbation column is not a transfer result.
Swipe horizontally to read the table
Model Perturbation Raw Perturbation Checked Injection flip Injection target Jev (construction model) 69.7 36.6 54.4 49.6 KEV 49.1 20.0 22.5 18.3 AutoTrust 48.8 12.5 64.6 60.7 OpenJEV 45.4 12.8 86.3 82.5 Bosun 38.9 22.5 47.9 41.1 Liquid D1 51.9 19.5 28.7 23.2 - Transfer and label checking: The authors state attacks constructed on Jev also succeed on other models, but Checked ASR generally drops. On Arm's Noul, the five transfer targets' Checked ASR is KEV 0.0, AutoTrust 3.4, OpenJEV 3.7, Bosun 5.9, Liquid D1 7.3 (Table 3).
- Task differences: The authors state that among transfer targets, OpenJEV has the highest Checked ASR on PromptShield's Noul at 36.0, and the lowest on MMLU's Noul at 13.5 (Table 3). On BFCL's Choice, AutoTrust, OpenJEV, and Bosun have Checked ASR of 0.0.
- Injection location: The authors state no single template is strongest across all models, locations, and metrics. Observation injection has higher flip and target ASR on Jev, KEV, AutoTrust, and Bosun; OpenJEV and Liquid D1 are the opposite, with instruction Combined at 78.0/74.8 and 18.6/15.7 (Table 4, Table A4). The authors state target ASR is always lower than flip ASR.
Model Differences on the Curated Set
- What was measured: Candidate-level perturbation ASR and injection target ASR are reported on 30 text models; candidate-level perturbation ASR, plus in-image injection and instruction injection target ASR, are reported on 17 visual models. The denominator is retained candidates where clean predictions are correct and attack responses are valid.
- Numbers named in the text: The authors state all held-out text models show errors in Figure 2. GLiNER2.5-Decide-340M has perturbation ASR of 11.4% and injection target ASR of 55.3%. In Figure 3, Strands Decider-2B has held-out image perturbation ASR of 31.8% and in-image injection target ASR of 81.2%; Perplexity Decider v1.1-27B has perturbation ASR of 3.2% and instruction injection target ASR of 45.7%.
- Author interpretation: Robustness to perturbation need not extend to injection, and the more effective injection location also varies by model. The authors state StandardOne-8B has the highest Checked ASR for common image perturbations across four datasets, and the lowest on FinQA at 23.8% (Table 5); overall ASR can mask task differences.
Image Perturbations and Injections
- Common perturbations: After merging five types and five levels, full-dataset Raw/Checked ASR is Imajev-4B 47.2/26.4, Visual-Jev-4B 50.9/30.6, Jev-Omni 51.1/33.2, OneJev-4B 49.8/25.4, StandardOne-8B 70.4/47.0 (Table 5, N=320). The authors state label checking lowers ASR on every dataset. Appendix A.4 states defocus blur remains the strongest type after checking, and merging types raises label-consistent ASR for all five models.
- Optimized perturbations: Table 6 has one image per Choice question, ϵ=8/255, N=160. Those marked with † are optimization surrogates: Imajev-4B 77.7/62.8, Visual-Jev-4B 82.8/64.5, Jev-Omni 72.3/56.3. Transfer targets across the full dataset are OneJev-4B 39.8/32.4, StandardOne-8B 7.0/6.2, StandardOne-3B 10.9/8.9, NeoHorse-Jev-4B 20.0/11.7. All four targets are 0.0 on Geometry3K. The authors state OneJev-4B is overall the most susceptible, but on Multimodal-Mind2Web, StandardOne-3B and NeoHorse-Jev-4B have Checked ASR of 40.0 and 25.0, higher than OneJev-4B's 18.2.
- Pre-filtering injections: Table 7 covers five visual filtering models with Choice inputs. The highest in-image Checked target ASR is Visual-Jev-4B at 79.4; Jev-Omni's instruction injection flip/target is 93.0/93.0, and it is the only one of the five where instruction injection is higher than in-image injection. The authors state label checking lowers in-image ASR for all five models, but injections can still change decisions when visible evidence still supports the reference answer.
Other Ablations and Analyses
- Table 3: On NevIR's Choice, Jev is 98.0/20.4 and Liquid D1 is 95.2/19.0; on Siim's Noul, Jev is 85.9/19.7 and AutoTrust is 74.3/5.7. The authors state most raw successes at these two locations fail label checking.
- Table A1: Across 16 source groups, combined attacks have higher raw ASR than Repeat@8, and higher Checked ASR than Paraphrase@4. On SST-2 (N=98), Jev's Repeat@8 is 0.0, Para@4 is 2.0/1.0, and Combined is 96.9/70.4.
- Table A2: On four shared datasets with 488 clean-correct samples per method, average queries are Charmer 5,755.7, SememePSO 2,027.4, TextFooler 329.4, DeepWordBug 115.3. The paper states this comparison does not control the query budget.
- Table A3: On the external check pool, Charmer's ASR is Jev 93.1, KEV 28.2, OpenJEV 14.2, Bosun 20.4. SCPN is 12.5 on Bosun, higher than Jev's 5.4. The authors state effects on the construction model do not fully predict other models.
- Table A7: After changing Jev's decision format, full-dataset Choice to Noul goes from 28.7 to 20.3 (N=916), and Noul to Choice from 18.3 to 13.4 (N=883). The authors state every dataset drops but remains nonzero; PromptShield drops the most and SST-2 the least. The two candidate sets differ, so the authors believe the two formats cannot be compared this way.
- Tables A8 and A9/A10: After switching the same 160 image pairs to Noul, full-dataset raw ASR goes from 82.8 to 10.8 for Visual-Jev-4B and from 77.7 to 65.3 for Imajev-4B. The authors state Visual-Jev-4B, though in the surrogate set, is much lower on Noul. On both text sets, accuracy and reference probability drop across all 30 models; on injections, OpenJEV's ∆p is -68.2 and KEV's is -12.1 (N=374, Np=318). On image perturbations, GEV-26B accuracy goes from 78.0 to 80.0 with ∆p of -0.3, the accuracy exception the authors note.
What are some points that can be further explored?
The authors note that the benchmark only looks at single decisions, image attack transfer is limited, and computational constraints also limit method and task coverage.
Limitations and Future Directions Noted by the Authors
- Single decisions only: The benchmark focuses on individual decisions, not the downstream impact of these decisions in full workflows (Limitations).
- Image transfer: Image perturbations transfer limitedly to some held-out models (Limitations).
- Multi-agent optimization: Optimizing against multiple surrogates does not guarantee broad transfer (Limitations).
- Computational constraints: Computational constraints limit the attack methods and configurations evaluated, as well as their coverage across tasks (Limitations).
- Future directions: The authors propose exploring more transferable image attacks and expanding the attack suite (Limitations).
Experimental Coverage
- The curated evaluation covers 30 text model configurations and 17 visual model configurations, of which 24 and 12 did not participate in candidate filtering (Section 3, Figures 2–3).
- Before filtering, the text perturbation pool is 10,093 candidates, 1,332 samples, 16 source groups; the injection pool is 9,716 candidates, 1,232 samples, 15 source groups, with PromptShield excluded (Section 3.2).
- The curated set is 141 text perturbations (120 samples), 374 injections (265 samples); 200 image perturbations (127 common, 73 optimized), 223 in-image injections, 120 instruction injections (Sections 3.2.3, 3.3.3).
- Label checking uses GPT-5 mini, DeepSeek-V4-Flash, Claude Sonnet 5, and GPT-5; text perturbations additionally undergo GPT-5 semantic review and per-case human review, and common image perturbations undergo human readability checks (Section 2.3.3). The paper does not report repetition counts for the main evaluation, nor agreement numbers between external models and human review.
- Eight text editing methods are compared on average queries across SST-2, MMLU, PromptShield, and SMS Spam; the paper states this comparison does not control the query budget. Charmer and SememePSO are not used on some source groups due to computational constraints (Appendix A.1, Tables A1–A2).
Summarize the main content of the paper
DecisionRobust evaluates the robustness and attack transfer of JEV-style decision models under text and image perturbations and prompt injection.
DecisionRobust evaluates the adversarial robustness of JEV-style decision models on application tasks, covering text and images.
- Problem: Models receive instructions, options, or scoring criteria at inference time and map observations to Choice, Noul, or Score for routing, filtering, and approval. The authors believe observations and images may be attacker-controlled, and outputs that conform to the format can still select the wrong action; clean accuracy is insufficient.
- Approach: Perturbations aim to preserve task-relevant information and cause arbitrary errors; injections target predefined wrong answers. Text adaptive attacks use Jev's (jev-1.13.0) probabilities or scores without parameter or gradient access. Image optimization minimizes the sum of log probabilities of clean predictions across three surrogates. Candidates undergo label-consistency checks by external models and cross-model filtering; 24 text models and 12 visual models do not participate in filtering.
- Text numbers: Before filtering, across the full dataset, Jev's perturbation Combined ASR is raw 69.7% and checked 36.6%; OpenJEV's injection target ASR at Either location is 82.5% (Table 3, Table 4). The authors state Checked is generally lower than Raw, and no injection template is strongest across all models.
- Images and curated set: Common image perturbation Checked ASR across the full dataset ranges from 25.4% to 47.0% (Table 5). Optimized images have Checked ASR of 64.5% on surrogate Visual-Jev-4B and 6.2% on non-surrogate StandardOne-8B; all four transfer targets are 0.0 on Geometry3K (Table 6). On the curated set, GLiNER2.5-Decide-340M has perturbation ASR of 11.4% and injection target ASR of 55.3% (Figure 2).
- Author conclusions: Attacks cause errors beyond the construction and filtering models, but optimized image perturbations transfer unevenly. The authors believe robustness to perturbation does not extend to injection; label checking removes many apparent successes, so attacks must be verified as still effective.