acceptodds
Under review as a conference paper at ICLR 2027

BadTouch: The First Study of Data Poisoning in Tactile Models

Abstract

Repeated tactile observations share a physical contact, but their supplied force labels need not share the same quality. How should a curator use this provenance when adapting a frozen tactile encoder? We study label-only poisoning on 904 public GelSight trajectories with same-input controls, two frozen encoder settings, and explicit verified-label budgets. A finite-bank deletion frontier separates errors made by a detector from correct samples necessarily lost through whole-contact removal. In the Sparsh curation bank, complete poison removal by whole-contact deletion requires deleting 93.69% of correct sharp candidate rows. Using the same trusted predictor and contact-calibrated thresholds, row deletion reduces this loss to 0.26% and sharp MAE from 0.092 to 0.054 N, at 99.67% poison recall for the 1.5 N target. This collateral improvement does not establish additional task value: when all 362 verified contacts can be reused, excluding unverified sharp candidates reaches ordinary/sharp MAE of 0.025/0.040 N, compared with 0.026/0.039 N after filtered augmentation. Lower-budget sharp gains vary across seeds. Source-allocation, calibration-scarcity, accepted-label and recovery controls delimit these findings within one acquisition population. The results show why provenance-aware curation must evaluate both deletion cost and predictive utility under the same available verified information.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
tactile sensing, data poisoning, backdoor attacks, robot learning, tactile foundation models, machine learning security, data curation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_nSGPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies label-only poisoning of a small MLP force-regression head on top of frozen tactile encoders (Sparsh and ImageNet DINO). It uses 904 public GelSight trajectories. An attacker relabels some sharp-indenter frames to 1.5 N or 5 N. With 48 poisoned frames (0.58%), sharp-condition MAE rises by up to 0.66 N while ordinary MAE barely changes. On the defense side, the paper notes that deleting whole contacts with complete poison recall costs 93.7% of correct sharp rows (Prop. 1). It shows that row-level deletion, using the same contact-calibrated conformal threshold, loses only 0.26%. When all verified labels are reused, simply excluding unverified sharp data does as well as filtering (0.025/0.040 vs. 0.026/0.039 N). Adaptive labels evade the screen, and fine-tuning on trusted data removes most of the effect.

Critical strengths

Separating deletion cost (correct data lost) from final predictive utility is a sensible evaluation principle for curation pipelines.

Critical weaknesses
  1. The threat model makes the defense problem trivial, and the paper's own result shows this. Condition metadata is trusted and poisons are confined to sharp candidates. So the curator can drop unverified sharp rows, and Table 2 shows this matches the filtered pipeline. The extra sharp gains are within seed noise (e.g., −0.0014 ± 0.0040 N). The adaptive attack evades the screen (29% rejected), and trusted fine-tuning undoes it. The study therefore does not show a realistic risk, and it does not show that the defense adds value.

  2. The technical content is limited and the title overclaims. Prop. 1 is immediate, as the paper says. Row deletion removes fewer correct rows than contact deletion by construction (Drow⊆DcontactD_{\text{row}}\subseteq D_{\text{contact}}). The calibration guarantee is standard split conformal. Poisoning a small regression head on frozen features is a well-known setting. Novelty comes only from the tactile domain: one dataset, one acquisition population, and offline force estimates with no controller or physical consequence. "BadTouch: The First Study of Data Poisoning in Tactile Models" does not match a paper whose conclusion is about curation accounting. The text is also very dense and heavily hedged, which makes the main message hard to recover.

Questions for authors
  1. Is there a threat model where category exclusion is not available (untrusted condition metadata, or poison spread across conditions) and your row-level screen still adds utility? Results there would change my assessment.

  2. Does the effect hold on a second tactile dataset or sensor (e.g., DIGIT, or AnyTouch data)?

  3. Does poisoning still matter when the encoder is fine-tuned rather than frozen, or when the head is trained with a robust loss (e.g., Huber, which you cite)?

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

This paper looks at a security weak spot in how robots learn to sense touch. Many systems use a fixed, pre-trained tactile encoder that turns images from a touch sensor (GelSight) into useful features, then train a small add-on model to estimate how hard something is pressing. That setup is cheap and convenient, but if some of the training data comes from an untrusted source, a bad actor can quietly teach the system to report the wrong force whenever it touches a particular kind of object. The authors simulate this by taking touches from a sharp indenter and swapping their true force labels for false ones, as believable value (1.5 N) or an obviously out-of-range one (5 N) that a basic check can catch. The paper enquire 3 questions for anyone cleaning up such data. First, how much do the corrupted labels actually change the model’s behaviour? Second, how much good data gets thrown away when you delete every recording of a suspicious contact, compared with deleting only individual suspicious readings? Third, does keeping the flagged data, after filtering, produce a better final model than simply retraining on data you already know is trustworthy? Apparently, the main take away is that keeping more data is not the same as getting a better model.

Critical strengths
  • The submission claims to be the first of its kind for data poisoning in tactile models.
Critical weaknesses
  • The readability of the submission is extremely poor for a human to read and understand it.
  • The submission heavily relies on AI usage, which makes it susceptible to potential hallucinations.
Questions for authors

None.

Reviewer_ibKkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies label-only data poisoning for force estimation with frozen tactile encoders. The authors corrupt labels from the "sharp" contact condition of the GelSight force dataset, and compare whole-contact versus row-level deletion using a trusted residual predictor with contact-calibrated thresholds. Experiments with Sparsh and ImageNet DINO encoders show that row deletion greatly reduces the loss of correctly labelled sharp candidates at similar poison recall. However, this retention advantage does not yield consistent predictive gains once all verified labels are reused by the final learner.

Critical strengths
  • Tactile perception is underexplored for poisoning research. The repeated observations from the same physical contact also make it easy to have grouped data.
  • The same-input, correct-label controls carefully isolate the effect of label corruption. Accounting for verified-label reuse is valuable and reveals limitations that a filtering-only evaluation would miss.
  • Data provenance, partitions and prediction records are provided. The authors also explicitly stated the limitations of the adaptive attacks.
Critical weaknesses
  • The novelty of this paper is limited. The attack follows established subpopulation poisoning [1], Proposition 1 is an immediate counting argument, and the row-deletion guarantee follows from standard split calibration and set inclusion. Evidence comes from one dataset and one force-estimation task, so it is unclear which findings are specific to tactile sensing rather than generic grouped regression.
  • A simple label-only baseline is missing. Tables 11-12 report an AUROC of 0.997 for the maximum-label score, yet the range audit uses a fixed 3 N threshold and therefore misses 1.5 N replacements. Moreover, 1.5 N already exceeds the observed sharp batch-1 range (max 1.33 N). An upper-tail label filter calibrated per condition/batch on the same verified contacts should be compared, as now the added detection value of tactile features remains unclear.
  • In the 48-poison experiments, every added sharp sample is corrupted, and in the curation experiment half of the unverified sharp candidates are poisoned. The 93.69% collateral also relies on poisons being spread across nearly all source contacts. The source-concentration and corruption-mixture experiments should be connected to the filtering and retraining pipeline, using the in-range 1.5 N target rather than the easily detected 5 N one.
  • The practical utility is not established. With equal verified-label access, excluding all unverified sharp candidates performs on par with filtered augmentation, and even the lowest budget includes 42 verified sharp contacts. With scarce calibration data, the threshold becomes infinite and recall drops to zero. The paper does not identify a regime where screening consistently outperforms simpler uses of the trusted data.
  • The different studies change splits, contamination structure and verified-label usage, which makes them hard to follow. A consolidated protocol table would help, and BayesClean I1 should be defined in the main text before Table 3.

[1] Jagielski et al. Subpopulation Data Poisoning Attacks. ACM CCS 2021.

Questions for authors
  • Does the 1.5 N attack remain effective with substantial correct sharp coverage and low within-sharp contamination? How does source concentration change the row-versus-contact deletion comparison?
  • When a contact contains both high-residual and low-residual poisons, whole-contact deletion removes poisons that row deletion retains. How do the two policies compare on the recall–retention tradeoff in this case?
  • When sharp fitting and calibration labels are jointly scarce, is there a setting where filtering consistently outperforms direct training on the same trusted-label budget?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.