acceptodds
Under review as a conference paper at ICLR 2027

Row Selection in Context-Label Attacks on Tabular Foundation Models

Abstract

A tabular foundation model uses labeled context rows to predict new rows. Changing a small number of context labels can therefore affect many predictions. We study how the choice of rows changes this damage. We compare neighbor-degree, mean-distance, and query-coverage selectors across TabPFNv2-12L, TabICLv2, and Mitra. Class-matched random controls separate total corruption damage from the additional damage due to row selection. Hub attacks span five count budgets and four percentage budgets. Their total loss grows with corruption, while additional damage depends on geometry and budget. On eight external datasets, representation hubs and random selection have equivalent normalized loss at 25 labels within a margin of 0.005. At 15%, input hubs and mean-distance selection add 4.16 and 2.60 percentage points of accuracy loss relative to their respective controls. Input anti-hubs, which appear in few neighbor sets, add 1.00 percentage point at 5% and retain additional loss under coverage matching, repeated tie choices, and row permutations. In an exploratory single-label analysis, query coverage has a clearer association with influence than neighbor degree. Selecting high-coverage rows adds 1.38 percentage points of accuracy loss at 5%. Together, these results identify how selection geometry, attack size, and the prediction measure shape context-label attacks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
tabular foundation models, in-context learning, context-label poisoning, adversarial robustness, hubness

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RK42OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how selecting particular context rows affects the damage caused by label corruption in tabular foundation models. It evaluates input-space and representation-space hubs, input anti-hubs, mean-distance selection, and query-coverage selection on TabPFNv2-12L, TabICLv2, and Mitra. The experiments show that selection effects depend on the attack budget, geometry, and evaluation metric. Representation-hub selection is practically equivalent to matched random selection in normalized loss at the specified 25-label setting. Input anti-hubs produce additional damage at modest budgets, including a 1.00 percentage-point additional accuracy drop at 5% corruption. Query coverage also provides an effective selection rule for some outcomes. The study includes three models, five splits, an eight-dataset main panel, broader hub experiments, and robustness checks for anti-hub selection.

Critical strengths
  1. The controls isolate a meaningful effect of row selection. Matching source-class counts and replacement labels avoids attributing differences in corruption volume or class composition to the selection rule. This is consequential because substantial total attack damage can coexist with little additional damage over matched random corruption. The paper makes that distinction explicit and evaluates it across budgets.
  2. The anti-hub result is supported by several targeted checks. At 5% corruption, anti-hubs add 1.00 percentage point of accuracy loss over matched random selection, with positive dataset-level accuracy effects on all eight main datasets. The authors also examine alternative tie choices, row permutations, and coverage-matched controls. These checks make the finding more convincing than a single comparison and challenge the intuitive expectation that highly connected rows should necessarily be the most damaging to corrupt.
Critical weaknesses
  1. The practical significance of the threat model needs clearer grounding. The manuscript specifies which information each selector uses, but provides limited discussion of a concrete deployment where an adversary can selectively rewrite context labels. Representation-based attacks additionally require access to internal representations, while query coverage requires access to query features. These assumptions may be reasonable in some settings, but they imply different attacker capabilities. A concrete example explaining how modified labels enter the inference context, and which selectors are feasible under that access, would make the security implications easier to assess.
  2. The generality of the main selection findings remains insufficiently established. The anti-hub and query-coverage results are evaluated primarily on eight datasets, while the broader 24-dataset experiments mainly investigate hubs. The geometric analysis fixes (k=10), and the representation-based selectors use one chosen layer per model. Sensitivity to neighborhood size, representation layer, and alternative multiclass replacement mappings is not established. Since the central conclusions concern geometry-dependent effects, these choices could affect the interpretation of both positive and near-zero findings. The existing tie and coverage checks are useful but address different sources of variation.
Questions for authors
  1. Robustness clarification: How sensitive are the anti-hub results to the neighborhood size (k), and how sensitive are the representation-hub conclusions to the chosen extraction layer? Were these settings selected before examining the main evaluation results?
  2. Evaluation clarification: How was the eight-dataset main panel selected, and was it held separate from selector development? Do the anti-hub findings persist on the remaining datasets in the broader panel?
Reviewer_3sonOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper studies how the choice of context rows affects label-poisoning attacks on TFMs. The authors use class matched random controls to isolate the precise marginal damage of targeted selection. The main results establish that attack effectiveness relies heavily on dataset geometry, corruption budget and the evaluation metric.

Critical strengths
  • The research question is relevant, as TFMs rely directly on labeled context examples at inference time. Understanding which rows are more vulnerable to label corruption is an interesting problem.
  • I appreciate the use of class-matched random controls, which makes it easier to distinguish the effect of row selection from general label corruption.
  • The anti-hub finding is interesting and somewhat surprising. The authors also make a good effort to validate it across different models and datasets with additional robustness checks.
Critical weaknesses
  1. The authors demonstrate that selecting anti hubs causes more damage than selecting hubs or random rows, but why this actually happens remains unclear. The exploratory single label analysis shows almost no association between anti hubness and individual label influence, yet the multi label attacks show a pronounced effect. Although the paper considers several alternative explanations, it misses a direct analysis of what actually makes these specific rows influential.

  2. The sensitivity to context size and neighborhood geometry is not sufficiently explored. The main experiments rely on context sizes between 400 and 1400 rows with a fixed neighborhood parameter k = 10. Since hubness inherently depends on local context structure, it is unclear whether the findings hold under much smaller or larger context windows. The effect could also shift depending on the choice of k, the distance metric, or the representation layer. A thorough sensitivity analysis is necessary to establish how general these observations actually are.

  3. Related work, such as [1] has already studied geometric label-poisoning attacks and compared TFMs with conventional classifiers. I think a similar comparison here, particularly with kNN or tree-based models would be useful. If anti-hub selection is also effective for conventional classifiers, the finding may reflect a more general property of tabular label corruption rather than something specific to TFM inference.

[1 ]Biloš et al., "A Mechanistic Study of Tabular Foundation Models."
Questions for authors

Q1. Why do anti-hubs cause more damage in multi-label attacks when anti-hubness shows little association with single-label influence? Could interactions between corrupted labels explain this?

Q2. How sensitive are the results to context size and the choice of (k)? Would the anti-hub effect still hold with substantially smaller or larger contexts?

Q3. Have you considered comparing anti-hub attacks with conventional classifiers, such as kNN or XGBoost? I think this would help clarify whether the effect is specific to TFMs or a more general property of tabular data.

Overall, I find the anti-hub observation interesting and appreciate the careful experimental comparisons. However, I am not yet fully convinced that the paper provides enough understanding of why this effect occurs or how broadly it applies beyond the current experimental setting. I would be happy to reconsider my score if the authors can adequately address these concerns, particularly Q1 and Q3.

Reviewer_i75fOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies which labeled context rows an attacker should alter to damage a tabular foundation model's predictions. It compares rows that appear in many nearest-neighbor sets, rows that appear in few such sets, mean-distance selection, and selection based on proximity to query rows. Each attack is compared with random relabeling that preserves the number of changes from each original class and their replacement labels. This isolates the additional damage caused by choosing particular rows.

Experiments use TabPFNv2-12L, TabICLv2, and Mitra 1.1, with eight datasets for the main comparisons and a broader panel for budget curves. At a 5% corruption budget, anti-hubs cause an additional 1.00 percentage point of accuracy loss over matched random and retain an advantage under coverage matching and tie-handling checks. The results show that the benefit of row selection depends on the geometry, corruption budget, and prediction metric.

Critical strengths
  1. The controls isolate a meaningful effect of row selection. Matching both source-class counts and replacement labels separates the damage from selecting specific rows from ordinary label corruption. Input hubs and anti-hubs also share their class flow and random controls, enabling a direct comparison. This design is the paper's strongest contribution because total attack damage alone cannot establish that a geometric selector helps.

  2. The anti-hub accuracy result survives several relevant checks. At a 5% corruption budget, anti-hubs add 1.00 percentage point of accuracy loss over matched random, with a 95% interval of 0.43 to 1.65 and positive dataset-level differences on all eight datasets. The effect is positive for each of the three model families, exceeds input hubs by 0.63 points, and retains 0.91 points under coverage matching. Repeated tie choices and row permutations provide further support for the result.

Critical weaknesses
  1. The practical-equivalence margin is weakly motivated. Representation-hub equivalence is established only for the averaged normalized loss at 25 changed labels within a margin of 0.005. The eight-dataset interval extends to −0.00457, close to that boundary. For scale, the margin is about 64% of the anti-hub effect of 0.0078 at 5% corruption. It therefore permits differences comparable with a substantial fraction of another reported attack effect. This limits the interpretation of equivalence even though the stated interval criterion is satisfied.

  2. The generality of the geometric conclusions remains uncertain. The main positive comparisons use eight datasets, one neighborhood size of 10, and one representation layer per model. In input space, a one-IQR numeric difference contributes 0.1 while a categorical mismatch contributes 1, making the selected rows sensitive to the distance construction. The paper also does not explain what makes the test panel external or which choices preceded its evaluation. Query coverage's positive accuracy result should remain separate from its inconclusive primary normalized-loss result, whose 95% interval is −0.0038 to 0.0550.

Questions for authors
  1. What practical criterion motivates the 0.005 normalized-loss margin, and which equivalence conclusions survive smaller margins? Please relate it to clean loss and the attack effects, rather than only to uniform-prediction loss.

  2. What is the eight-dataset panel external to, and which hypotheses, selector settings, and margins were fixed before evaluating it? Can you provide the exact model configurations and representation-extraction procedure, including how the claimed absence of query access is enforced?

  3. Do the anti-hub accuracy advantage and representation-hub equivalence persist for other neighborhood sizes, numeric and categorical weightings, and nearby representation layers? A compact check would materially strengthen the claim that the observed behavior extends beyond the chosen geometry.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.