acceptodds
Under review as a conference paper at ICLR 2027

SensorAD: Cross-Modal Physical Consistency for Sensor Attack Detection in Autonomous Vehicle

Abstract

Autonomous vehicles fuse cameras, LiDAR, event cameras and GPS/IMU to perceive and act safely. Each of these sensors is individually attackable, but an adversary who compromises one cannot easily forge physically consistent signals across the rest. SensorAD turns this asymmetry into a detector: on normal driving data we mask an entire sensor modality and train a unified transformer to predict it from the remaining ones, so that the error minimized in training with a modality masked is the score read at deployment and no attack labels are used for training or threshold calibration. We pre-train on real-world nuScenes data, fine-tune on CARLA to incorporate the event camera, and evaluate on a 132-episode benchmark spanning ten attack types along six axes. Against nine baselines from five paradigms, SensorAD is strongest on physics-grounded attacks, averaging 0.843 AUROC versus 0.738 for the best cross-modal baseline; it localizes the compromised sensor at 67.0% top-1 / 79.5% top-2 accuracy, and its timing advantage over that baseline is specific to stealthy attacks: it flags a gradual GPS drift after 3.4±0.4 m of accumulated offset against 12.6±1.4 m. Detection degrades gradually under coordinated, spatially aligned and white-box adaptive attackers and costs at most 0.028 AUROC on unseen towns and weather. On real nuScenes streams with the same threat models injected, the CARLA-fine-tuned detector reaches 0.735 mean AUROC on the five expressible attacks (0.793 in CARLA), though its threshold must be recalibrated on real data. Per-sensor visual detectors remain stronger on weakly cross-modal attacks: the two families fail on largely disjoint attack classes, and a simple late-fusion stack outperforms either alone on average.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Autonomous vehicles, sensor attack detection, multimodal anomaly detection, cross-modal consistency, self-supervised learning, adversarial robustness

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4WboOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes to train a model to be robust even when one of its modality is masked. It is tested on nuScene and CARLA to demonstrate the performance of AV when using this newly trained model.

Critical strengths

The idea of training a model to be robust even when one of its modality is masked is not novel, but interesting. Especially, the authors wanted to look at the effect of masking on the overall system, i.e., does it affect safety.

Critical weaknesses

The paper is VERY hard to read and follow. A complete rewrite is warranted.

Related work: explain differences with • Chen, Liu, and Pei, “Cross-Modal Matching CNN for Autonomous Driving Sensor Data Monitoring,” ICCV Workshops 2021. This work explicitly assumes that overlapping camera and LiDAR observations should be consistent under normal operation and detects anomalies, including sensor mis-synchronization, by matching the two modalities. It also evaluates sensor fault detection and point-cloud quality assessment on KITTI. • Cao, Liu, and Li, “Leveraging Cross-Modal Consistency for Sensor Fault Detection in Autonomous Driving,” IEEE IV 2026. This paper explicitly models cross-modal consistency among cameras and LiDAR, trains modality-specific VAEs on fault-free data, and uses deviations in the joint latent space to detect and classify sensor faults.

Use of AI: - The long sentence " This is a design correspondence, not a theorem. Note where the shift enters: the masked pass sees [MASK] in place of modality k plus the uncompromised sensors, exactly its training input, so it is not shifted at deployment; the reference pass producing the target does see the corrupted tokens, which is why a score trained only on normal data can flag attacked frames at all." seems to be 100% from AI. "[MASK]" has never been specified. - Clearly the analysis (see page 7) is all written by AI. - If you are honest with yourself, you cannot understand what is written by AI. You have a responsibility to write for other humans to understand. I appreciate your "AI Use statement" but read the paper please.

Need serious improvement in writing • I would up Figure 1 as this is core of the paper. I would also improve the writing by changing the structure and ordering of the paper. Until page 5, the reader has no idea what is happening. This reads as complicated sentences for no good reason. • What does "0.9.15 episodes" mean? • "Nine methods across five paradigms" the reader doesn't even know what methods is referred to here. The writing is very clunky and not intuitive. • Attacks aren't described beside the threat model give in Section 3, which states the level of access the attacker has. We need a proper Attacker model section with the detail on what exactly is modified by the attacker. Section 4.1 has the FPR@95%TPR and AUROC per attack, but we don't know what are the attacks. Later we see "physics-grounded" and "weakly cross-modal" but never described! • You shouldn't have content in Appendix that are "the most important". For example, Section 4.3, you mentioned about Table 4 that is in Appendix B.

Random comments:

  • "Problem formulation" is not a contribution
  • Line 138 what is DVS? Later in Figure 1, we see that this is linked to event camera, but at no point do you explain it. You need to assume folks aren't familiar with the domain.
  • From Proposition 1, I don't understand what you are trying to do here. Are you training a model to get it to predict what the "removed modality" would see, or the error it would have? How do you do that? How much information about the modality do you need? Honestly I would remove Section 3.1.
  • Section 4.6, what are the safety impact e.g., collision rate?
Questions for authors

I do not have questions that answers would materially affect my assessment. Just need a full rewrite.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 4Wbo

We thank the reviewer for their time. The review's summary describes a different method, so we first restate what SensorAD does and then answer each point. Line numbers refer to the submitted version; since the current PDF is the revision, we also give its locations.

What SensorAD is. SensorAD is not trained to be robust to a missing modality. It is an attack detector: each modality is masked so it can be predicted from the others, and the prediction error is the anomaly score (Eq. 2). This is stated in the abstract (l. 016–019) and on page 2 (l. 063–066).

Each point, and what the revision now does:

Review pointIn the submissionRevision
DVS not explainedl. 039–040, "dynamic vision sensors (DVS)"; Fig. 1explained at first use (l. 040–041)
[MASK] not specifiedFig. 1 legend; App. F; App. F.3defined in Sec. 3.1
"0.9.15 episodes"CARLA version (l. 302)"simulator version 0.9.15" (l. 319)
Nine methods not identifiednamed by paradigm, l. 305–311"listed here by paradigm" (l. 333)
Attacks not describedTable 26 (l. 304); Table 1 captionnew Table R1 in Sec. 4.1
Physics-grounded / weakly cross-modalSec. 4.2, l. 343–346defined in Sec. 4.1 next to Table R1
Key results in appendixTable 4 in App. BTable 4 moved to Sec. 4.3
Hard to follow until page 5—plain-language overview opening Sec. 3

Proposition 1: what is predicted, and how much information is needed. The model predicts the hidden modality's features from the other sensors; the error is the score. At test time nothing from the hidden modality is used: one learned [MASK] token replaces it. Proposition 1 states what is needed instead: an attack is separable in expectation once its feature-space perturbation exceeds 2√ε_normal, where ε_normal is the remaining sensors' prediction error on normal data. Better cross-modal prediction therefore lowers the size of detectable attacks.

The quoted sentence (l. 179–182), in plain terms: with one sensor hidden, the model sees the same kind of input as in training, so its prediction stays reliable during an attack; the attacked sensor's data is used only as the comparison target, so an attack appears as a large gap.

Attacker model. Section 3.1 gives the attacker's access (Definition 2, l. 140–142; "Threat model", l. 147–151), and Table 26 gives exactly what each attack modifies; Table R1 now summarizes it in the main text.

Safety impact. Section 4.6 and Table 19 quantify each attack's damage (e.g., 18.3 m ego-pose error under GPS spoofing; −8.7 3D mAP under LiDAR tampering). Collision rate depends on the planner and controller, outside the scope of a sensor-level detector.

Related work. Like SensorAD, Chen et al. (2021) and Cao et al. (IV 2026) use cross-modal consistency to detect sensor faults on KITTI. Neither addresses adversarial attacks, GPS/IMU ego-motion consistency, leave-one-modality-out prediction, or label-free sensor attribution. Both are now cited and distinguished in Section 2.

On authorship. Our AI Use Statement discloses LLM assistance for copy-editing and for coding support; LLMs were not used to generate the idea, design the method or protocol, or produce or analyse results. Remarks about the authors' honesty and competence are not an assessment of the work.

We have raised our remaining concerns with the AC.

Thank you for providing additional details. I still believe that the Section 3.1, paragraph "Threat model" does not detail the attacks. If this is an attack paper, you must describe it fully. See how this is done in security conferences if needed. Re related work, could you demonstrate how the related work wouldn't be effective against the attacks?

Both points are already addressed in the revised PDF.

Attacks. The revision added Table R1 to Section 4.1 (main text), which states what each attack modifies. Table 26 gives exact parameters and real-world references. Section 3.1 now also states the attacker's goal and lists every attack:

  • GPS spoofing, drift and constant offset (up to ≈111 m);
  • 500 phantom LiDAR points;
  • RGB and DVS noise;
  • depth rescaling;
  • RGB+LiDAR denial of service;
  • replay of clean recordings;
  • a white-box GPS attacker (App. D.5).

The access model is in Definition 2 and the threat-model paragraph: the attacker can replace one sensor's output arbitrarily. SensorAD is a defense, and these attacks are its evaluation.

Related work. The revision already distinguishes both works in Section 2. Both monitor camera–LiDAR consistency only, for benign faults. Four of our eight attacks (GPS spoofing, drift, static offset and DVS noise) alter sensors those methods do not observe, so they cannot detect them by construction. Restricting SensorAD itself to camera and LiDAR lowers its overall AUROC across all attacks from 0.737 to 0.644 (Table 23). Our coordinated, spatially aligned and white-box attackers (Sec. 4.4, Table 7) target exactly the cross-sensor consistency such methods depend on.

Reviewer_H9CCOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors proposed sensorAD, an attack agnostic defense leveraging cross-modal physical consistency. The authors presented comprehensive evaluation in terms of how the proposed mehtods perform against different attacks and demonstrated strong performance against physcis grounded attacks.

Critical strengths
  • The proposed method follows a clear intuition where cross sensor consistency exists and is physical grounded. With that principle, the authors propose an attack agnostic framework to detecting anomalies.
  • The authors conducted comprehensive experiments to validate the hypothesis and demonstrate the effectiveness of the proposed method.
  • The authors disclose the limitation of the method and conducted detailed analysis for the usage of the proposed attack detection framework.
Critical weaknesses
  • Hypothesis on physics grounding across all modalities might be too strong. For example, camera might share more similar physics information with LiDAR but less with GPS
  • Sensor attacks against multi-modal also exist so the assumption of only one modality being compromised might not be true
  • Gap still remains in order to deploy, e.g. latency, threshold settings, additional mechanisms needed for detecting all suites of attack
Questions for authors
  • Can the detector distinguish attacks from benign sensor disagreement? Weather, calibration drift, timing errors, and sensor faults can also break cross-modal consistency.
  • How much real-world recalibration is needed and how robust it is to new enviroments?
  • How robust are results to stronger consistency-preserving attackers?
1 comment
AuthorsOpenReview ↗

Response to Reviewer H9CC

We thank the reviewer for recognizing the clear intuition, the comprehensive experiments and the detailed limitation analysis. We answer each weakness and the three questions listed in the ethics field, including new experiments.

W1. Physics grounding may not hold equally across all modalities. We agree, and SensorAD does not assume it does. Each modality is predicted from all others through learned attention, so the model learns how strongly each pair is coupled. The results reflect this:

  • LiDAR + GPS/IMU is the strongest two-modality pair (0.711 vs. 0.644 for RGB + LiDAR and 0.648 for GPS/IMU + DVS; Table 23), where geometry meets ego-motion.
  • Depth tampering surfaces mainly as disagreement with LiDAR, the other geometric sensor (68.4%, Table 5).
  • Appearance-only attacks carry little cross-modal signal and are where SensorAD is weakest, which is why we report the physics-grounded / weakly cross-modal split and recommend late fusion with a per-sensor detector (App. C.2).

The revision states the hypothesis as applying where physical coupling exists (Sec. 1), not uniformly.

W2. Multi-modal attacks exist. We agree; Section 4.4 evaluates them (Table 7):

SettingCFMSensorAD
2 of 4 sensors (RGB + LiDAR)0.665 ± 0.0240.717 ± 0.065
3 of 4 sensors (RGB + LiDAR + DVS)0.602 ± 0.0280.641 ± 0.026
Aligned RGB patch + LiDAR frustum0.611 ± 0.0270.653 ± 0.031

On average, detection degrades gracefully and stays above the cross-modal baseline, because the intact sensors still constrain the prediction. The necessary condition is that at least one intact sensor is physically coupled to the attacked ones; this limit applies to any consistency-based detector. A fully consistent attack rendered from a shared 3D object (Tu et al., 2022; Cao et al., S&P 2021) is strictly harder and is stated as open.

W3. Deployment gaps. We agree, and we make no deployment claim (App. G.5).

  • Latency: the single-pass score (45.8 ms, 21.8 FPS) fits the 50 ms frame budget at a small cost (0.724 vs. 0.737 AUROC, Table 14). The five-pass attribution runs only on flagged frames.
  • Thresholds: split-conformal calibration controls the realized FPR (App. C.1); real-data recalibration needs ~8 minutes of normal driving (Q2).
  • Full attack coverage: late fusion with a per-sensor detector lifts the eight-attack average from 0.722 to 0.750 (App. C.2).

Q1. Attacks vs. benign sensor disagreement (new experiment). We injected three benign faults into the 36 attack-free episodes over the same 1,000-frame window used for attacks, scored at the same calibrated threshold:

ConditionMagnitudeFrames flagged
Normal driving (Table 21)—1.4 ± 0.3%
Sensor dropout10% of LiDAR frames dropped8.9 ± 1.3%
Timestamp offset100 ms LiDAR delay14.8 ± 2.6%
Calibration drift1° yaw + 10 cm, ramped21.7 ± 3.2%
Attacks, same threshold—42.1%

Faults are flagged well above the normal rate because they break the same consistency attacks do. Their rates are below the 42.1% attack average, but that is an average, and we do not claim the score alone separates faults from attacks. For a safety monitor we regard flagging them as correct: a miscalibrated or desynchronized sensor is unsafe whatever its cause. SensorAD localizes the sensor to inspect and is meant to run alongside standard diagnostics (time-sync and calibration checks), which can establish a benign cause. The revision states that consistency cannot determine intent. Under weather shift SensorAD has the lowest FPR of all methods (Table 21), but shift still raises it (8.7% vs. 1.4%), so we treat environment-conditioned thresholds as a requirement.

Q2. Real-world recalibration (new experiment). Recalibrating on 1,000 real nuScenes keyframes (~8 min at the 2 Hz keyframe rate), using normal data only, brings the false-alarm rate to the 1% target, from 5.6% with the CARLA threshold. Only the threshold changes: ranking transfers (0.735 vs. 0.793 AUROC in CARLA, Table 11). Unseen towns and weather cost at most 0.028 AUROC but raise FPR at the original threshold to 2.6–5.8% (0.9–1.8% seen; Table 18), so each new environment needs recalibration.

Q3. Stronger consistency-preserving attackers. A white-box attacker optimizing the GPS trajectory against SensorAD's own score reduces AUROC from 0.851 to 0.792 / 0.711 / 0.632 at 0.5× / 1× / 2× budget (App. D.5). We claim no certified robustness. Replay attacks, locally plausible by construction, are detected at 0.748 (LiDAR) and 0.807 (GPS) (see the response to juhX).

Changes in the paper: hypothesis scoped to coupled pairs (Sec. 1); benign faults and the intent limitation (App. D.9, Table R3); recalibration cost and unseen-environment FPR (Sec. 4.5, App. C.3, Table 18).

We would be grateful if the reviewer would consider raising the score.

Reviewer_juhXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

SensorAD detects sensor attacks on autonomous vehicles by predicting each modality from the others and flagging disagreement. It uses modality masking, a shared transformer, and motion-related losses. It is pretrained on normal nuScenes data and fine-tuned in CARLA. Superior results are reported when compared with prior results (Crossmodal Feature Mapping (CFM; Costanzino et al., 2024)).

Critical strengths
  1. Cross-modal prediction is well motivated for locally plausible sensor attacks.

  2. Ablations and temporal-only controls support the benefit on GPS drift.

  3. The evaluation covers multiple seeds, environmental shifts, coordinated attacks, and real-stream injections.

  4. Limitations are discussed candidly.

Critical weaknesses
  1. Attacker capability exceeds the evaluation. The threat model permits arbitrary replacement of one sensor’s output, but most tests use predefined corruptions. Targeted replacement streams designed to cause harm while evading detection are largely untested; adaptive GPS attacks cover only a restricted family.

  2. Physical validation is limited. Attacks are digitally injected, including on nuScenes. These experiments do not validate physically realized spoofing or camera attacks.

  3. Baseline fairness remains uncertain. SensorAD uses dense tokens while CFM uses pooled features. Per-sensor DoS results also suffer from a documented normalization artifact.

  4. Generalization is incompletely established. Model selection uses the same attack families as evaluation.

Operational performance is weak. Recall is 42.1% at a 1% target FPR, false alarms increase under environmental shifts, and full scoring runs below the simulated sensor rate.

Questions for authors
  1. Does SensorAD withstand targeted replacement attacks within its stated threat model?

  2. Do gains survive corrected, representation-matched baselines and attack-family holdout?

  3. What are the alarm frequency and detection delay when computation is included?

2 comments
AuthorsOpenReview ↗

Response to Reviewer juhX

We thank the reviewer for a careful, specific review. We read the soundness concerns as (i) the threat model allowing more than the evaluation tests (W1) and (ii) model selection possibly leaking attack information (W4). Both are now tested with new experiments, and every question is answered.

W1 / Q1. Targeted replacement attacks (new experiment). Our adaptive attacker (Sec. 4.4, App. D.5) is targeted and harm-seeking: it must move the ego position by a target displacement while minimizing SensorAD's peak score. AUROC falls from 0.851 to 0.792 / 0.711 / 0.632 (TPR@5%FPR 0.648 → 0.557 / 0.432 / 0.318) across nested budgets.

To cover arbitrary replacement, we added replay attacks: one sensor's stream is replaced by a recorded clean stream from another route; the GPS replay starts at the true position, so there is no jump.

AttackAUROC
LiDAR phantom injection (Table 1)0.722 ± 0.070
LiDAR replay (new)0.748 ± 0.060
GPS spoofing (Table 1)0.823 ± 0.056
GPS replay (new)0.807 ± 0.051

A replayed stream passes any single-sensor check, but no longer agrees with the other sensors. Together the two experiments show the trade-off: to evade, a replacement must agree with what the intact sensors imply; to cause harm, it must depart from it. The white-box study fixes harm as a target displacement and measures how detection falls as the budget grows. It covers GPS, the modality an attacker shapes most freely; white-box LiDAR attacks remain untested.

W2. Physical validation. We claim no physical validation (Sec. 4.5, App. C.3, App. G.5). Our adversary controls a sensor data path (driver, ISP or bus), and injection at the sensor output is where that adversary acts. For a single attacked sensor this is also the harder case: digital injection gives full control of the stream, whereas a physical attack is further constrained (timing, angles, point counts). Parameters follow documented physical attacks (Table 26), and the threat models transfer to real nuScenes streams (0.735 mean AUROC, Table 11).

W3 / Q2a. Baseline fairness.

  • Representation matching: App. D.2 retrains SensorAD on CFM's pooled features. It still leads on physics-grounded attacks, 0.802 vs. 0.738, so (0.802 − 0.738)/(0.843 − 0.738) ≈ 61% of the margin comes from the cross-modal design.
  • DoS artifact: it stems from Anomalib's min-max normalization on all-zero input (App. I) and affects only the per-sensor baselines' DoS column, not CFM. Excluding DoS, SensorAD leads CFM 0.799 vs. 0.737 (p = 3.9×10⁻⁷, 36 episodes; Table 3).

W4 / Q2b. Attack-family holdout (new experiment). Training and threshold calibration use no attack labels. Model selection (architecture, loss weights, masking probabilities, modality subsets) did use labels, on a separate 30-episode pool disjoint from the benchmark (App. G.3), so no benchmark episode influenced any choice. Repeating the loss-weight selection for each family using only the other nine families and evaluating on the held-out one changes AUROC by −0.006 ± 0.008, within seed variation; the other choices were not re-selected per family.

W5. Operational performance.

  • Recall: 42.1% is frame-level. Per episode, with the 3-of-5 rule below, 73.0% of attack episodes are detected, while 15.7% of 100 s normal episodes raise an alarm. At α = 0.05, precision is 0.921 and F1 0.735 (Table 20).
  • Shift: lowest FPR of all methods in every shifted condition (Table 21), but shift still raises it (8.7% weather, 10.9% day→night; 1.4% normal), so environment-conditioned thresholds are required; real-data recalibration needs about 1,000 keyframes (~8 min).
  • Sensor rate: the single-pass score (45.8 ms) fits the 50 ms budget at a small cost (0.724 vs. 0.737 AUROC; 58.1% vs. 67.0% top-1 localization; Table 14). Five-pass attribution runs only on flagged frames.

Q3. Alarm frequency and delay with compute (new experiment). An alarm fires when ≥k of the last n frames exceed the α = 0.01 threshold; one hour = the 36 normal episodes:

RuleFalse alarms / hourAdded delayAttack episodes detectedNormal episodes alarmed
3-of-56.7 ± 2.52.7 ± 0.6 frames73.0%15.7%
4-of-52.7 ± 1.53.7 ± 0.6 frames69.7%7.4%
5-of-101.3 ± 1.55.7 ± 0.6 frames66.2%3.7%

Without a rule, ~1,000 frames/hour are flagged (1.4% at 20 Hz). With 3-of-5, end-to-end delay on abrupt attacks is ≈ 0.65 s: (9.4 + 2.7) frames at 50 ms (Table 8) plus 45.8 ms compute. Stricter rules trade detection and delay for fewer alarms (App. D.8, Table R2). Five-pass attribution (106.7 ms) runs on flagged frames only.

We would be grateful if the reviewer would consider raising the score.

AuthorsOpenReview ↗

Follow-up on responses to Reviewer juhX

We would be grateful to know whether our responses and the new experiments (replay attacks, attack-family holdout, persistence-rule alarm rates, and detection delay including computation) address your concerns. We are happy to clarify anything before the discussion period closes.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of responses and new results

We thank all reviewers. Reviewers found cross-modal prediction well motivated for locally plausible attacks (juhX, H9CC), the evaluation comprehensive across seeds, environment shift, coordinated attacks and real-stream injection (juhX, H9CC), and the limitations candidly stated (juhX, H9CC).

What SensorAD is. SensorAD is a label-free attack detector. On normal driving data it learns to predict each sensor from the others. At test time, a compromised sensor no longer agrees with the rest, and the prediction error both flags the attack and identifies the sensor (Eq. 2). The training objective and the deployment score are the same quantity, so no attack labels are needed for training or threshold calibration.

New experiments (3 seeds, mean ± std, same protocol as Table 1):

ExperimentResultAddresses
Replay attack: recorded clean LiDAR stream substitutedAUROC 0.748 ± 0.060juhX
Replay attack: recorded GPS trajectory, aligned to the true position at onsetAUROC 0.807 ± 0.051juhX
Loss-weight selection with the evaluated attack family held outΔAUROC −0.006 ± 0.008juhX
Benign faults, flag rate: calibration drift / timestamp offset / dropout21.7 ± 3.2% / 14.8 ± 2.6% / 8.9 ± 1.3%H9CC
Real-data recalibration to reach 1% FPR1,000 nuScenes keyframes (~8 min of driving)H9CC
3-of-5 persistence rule: false alarms on normal driving6.7 ± 2.5 per hourjuhX
3-of-5 persistence rule: added detection delay2.7 ± 0.6 frames (135 ms)juhX
FPR on unseen towns / weather, without recalibration2.6–5.8% (0.9–1.8% seen)H9CC

Main concerns and where they are answered. Threat-model coverage: replay and white-box attacks (juhX W1, H9CC Q3). Selection leakage: attack-family holdout (juhX W4). Operational performance: with a 3-of-5 persistence rule, 73.0% of attack episodes are detected while normal driving gives 6.7 false alarms per hour; stricter rules trade detection for fewer alarms (juhX W5, Q3). Benign disagreement and real-world thresholds: fault injection and recalibration cost (H9CC Q1, Q2).

Revision. The revised PDF marks every change in blue. It adds a plain-language overview at the start of Section 3, a compact attack table in Section 4.1 (Table R1), the ablation and localization table in the main text (Table 4), a discussion of Chen et al. (2021) and Cao et al. (2026) in Section 2, and all new experiments above (Sections 3.3, 4.4, 4.5; Appendices D.8–D.10 and E.3). New tables are numbered R1–R4, so every original table number cited in the reviews and in our responses is unchanged.

Dear Reviewers,

The rebuttal deadline is approaching. I am leaving this friendly reminder to encourage active discussion. Thank you for your valuable time and engagement!

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.