acceptodds
Withdrawn from ICLR 2027

Relational Injection Attack Against Relational Deep Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Relational deep learning (RDL) predicts outcomes for entities using information from database rows connected through foreign-key relationships. This reliance on relational context raises a question: can permitted row insertions manipulate an existing target’s prediction without modifying existing records or the trained model? We formulate Relational Injection Attack (RIA), a test-time attack that inserts linked rows under explicit write permissions. Under a white-box setting, our framework consists of two components: schema-admissible path selection and value optimization. The path selector ranks injection paths using the model’s response and the potential gains from attribute changes. For value optimization, we introduce Relational Projected Gradient Descent (RelPGD), which optimizes continuous weights over discrete candidate values using centered and scaled gradients, followed by multi-start local search on unsuccessful targets. We evaluate twelve binary classification tasks from seven RelBench datasets. Our path selector improves macro-average attack success rate by 10.1 percentage points over the relation-weight-norm heuristic. RelPGD matches or exceeds the mean attack success rates of CAPGD in 21 of 24 task–model settings. These findings demonstrate that permitted row injections can affect predictions, motivating their inclusion in RDL robustness evaluations. For transparency, our source code is available at [https://anonymous.4open.science/r/RIA-DCC0](https://anonymous.4open.science/r/RIA-DCC0).

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Relational Deep Learning, Graph Adversarial Attacks, Relational Databases, Graph Neural Networks, Tabular Adversarial Attack

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 4 reviews, above 27% of the others.
Reviewer_yqVaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies Relational Injection Attacks against relational deep learning models, where an attacker inserts new database rows while preserving existing records and model parameters. The formulation incorporates insertion permissions, referential integrity, temporal constraints, and budgets. The proposed framework combines path selection, RelPGD-based value optimization, and subsequent local search. Experiments cover twelve binary classification tasks from seven RelBench datasets and two victim architectures.

Critical strengths
  1. Meaningful database-specific setting. Requiring injections to be realizable through permitted database writes introduces substantive constraints beyond standard graph-level injection. The explicit treatment of permissions, foreign-key relationships, and temporal availability is a useful contribution to RDL robustness evaluation.

  2. Broad empirical coverage. The evaluation spans multiple datasets, tasks, and architectures, with component ablations and computational-cost analyses. These experiments provide useful evidence that valid row insertions can manipulate relational predictions.

Critical weaknesses
  1. Limited comparison with closely related injection attacks. Although the paper evaluates adapted tabular attacks, the path-selection baselines are mainly Random and W-norm. Relevant approaches—including G-NIA/OPTI (CIKM 2021), HeTa (IJCAI 2025), and TDGIA (KDD 2021)—are not experimentally compared. Their settings differ from RIA, so direct application may be inappropriate. Nevertheless, an adapted representative baseline under matched constraints and budgets, or a specific explanation of the adaptation difficulties, would help establish the methodological contribution beyond the valuable database setting.

  2. The headline path-selection gain is concentrated in one task. Table 1 reports a 10.1-percentage-point macro-average improvement over W-norm, but stack/user-badge alone improves by 90.4 points. Based on the rounded table values, the average improvement across the remaining eight tasks is approximately 0.075 points, with notable gains and losses on individual tasks. Further explanation and stability checks would strengthen the conclusion.

As a minor presentation concern, the first experimental results table appears only on page 8. Introducing a motivating empirical example or a compact results overview earlier would make the methodological development easier to follow.

Questions for authors
  1. Could the authors evaluate a representative adapted graph injection baseline, such as OPTI, under the same legal insertion space and computational budget? If this is impractical, which specific constraints prevent a meaningful adaptation?

  2. What explains the particularly large gain on stack/user-badge and the decreases on stack/user-engagement and event/user-repeat?

Reviewer_Rm5fOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies attacks that change a relational model's prediction by adding linked database rows while leaving existing records and model parameters unchanged. For example, adding a customer and a review could influence a product prediction. Each evaluated attack bundle contains a new dimension row and a fact row connecting it to the target. The proposed method ranks insertion paths using the model's response and gradient information, then applies RelPGD to optimize weights over candidate attribute values. Multi-start local search handles targets that the gradient attack does not flip.

The evaluation covers twelve binary RelBench tasks from seven databases and two victim models, HeteroSAGE and RelGNN. RelPGD matches or exceeds CAPGD's mean attack success rate in 21 of 24 settings before local search. Path-selection gains vary substantially by task and insertion budget, and local search removes most differences between these two gradient methods.

Critical strengths
  1. An attacker able to create a customer and review may influence a product prediction despite lacking permission to change the product itself. Representing this action as a linked pair of rows captures a constraint that unrestricted graph perturbations can miss. It is a meaningful direction for relational robustness research.

  2. Twelve binary tasks from seven databases are evaluated on HeteroSAGE and RelGNN. RelPGD matches or exceeds CAPGD in 21 of 24 settings in Table 10, including a gain from 37.63% to 52.96% attack success on RelGNN driver-dnf. Shared candidates and budgets make these optimizer comparisons informative, and unfavorable results are reported openly.

Critical weaknesses
  1. The defining validity claim cannot be assessed from the stated constraints. C1 to C4 describe general categories of permissions, foreign-key rules, timing restrictions, and budgets. They do not enumerate the allowed tables, columns, candidate sources, timestamp windows, or protections for label-defining observations in each task. A structurally valid insertion is not sufficient to establish that it is allowed under the declared attacker permissions. Because realizable database writes are the central distinction from graph injection, a per-task specification and decoded, validated example bundles are necessary to judge the contribution.

  2. The component gains do not establish a general advantage for the complete attack. At budget 2, the three-seed selector gain is 7.65 percentage points rather than the headline single-seed 10.1. Excluding user-badge, its average difference from W-norm is negative by 1.91 points. At budget 8, it trails W-norm by 1.17 points. Path selection is evaluated per target, while the main optimizer comparison fixes one majority-vote path per task. After local search, CAPGD+MLS and RelPGD+MLS have identical displayed means in 22 of 24 settings. These results support some useful component improvements but leave their combined benefit and total computational cost unresolved.

Questions for authors
  1. Can you provide a per-task manifest of writable tables and columns, candidate provenance, foreign-key and temporal constraints, and checks preserving label-defining observations? Please include decoded insertion bundles and explain how their validity is checked after rebuilding the graph.

  2. What are the attack success rate and total runtime of per-target path selection followed by RelPGD and local search, compared with matched alternatives on the same cohorts? Include selector calibration and final verification. How should the selector claim be scoped given the user-badge dependence and the reversal at budget 8?

  3. What are the clean AUROC, accuracy, and decision threshold for each victim and task? How much attack success is achieved by one random valid bundle per target, without search? The clean metrics clarify the operating points, and the additional control would distinguish optimizer gains from targets that are easily flipped by any insertion.

Reviewer_n1uVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Relational Injection Attack (RIA). Unlike conventional graph injection attacks that directly manipulate nodes and edges, RIA inserts new database rows while respecting schema, foreign-key, temporal, and insertion-budget constraints. The proposed framework consists of schema-admissible path selection and value optimization. The path selector combines insertion-induced prediction changes, gradient sensitivity, and estimated gains from admissible attribute substitutions. For value optimization, the authors propose Relational Projected Gradient Descent (RelPGD), which optimizes continuous weights over discrete candidate values using centered and scaled gradients, followed by multi-start local search for unsuccessful targets. The method is evaluated on 12 binary classification tasks from seven RelBench databases. The results demonstrates that schema-valid row insertions can manipulate predictions of frozen RDL models under a white-box threat model.

Critical strengths
  1. A relevant and clearly formulated attack surface for relational learning. The paper studies prediction manipulation through actual database row insertions rather than unrestricted changes to a graph representation. Explicitly accounting for foreign-key integrity, temporal availability, writable tables, and insertion budgets makes the problem formulation relevant to database-backed learning systems.

  2. Broad empirical evaluation analysis. The authors evaluate 12 classification tasks across seven RelBench databases and two RDL architectures. The experiments include comparisons with adapted constrained-attack baselines, path-selection ablations, insertion-budget variations, and computational-cost measurements. This provides useful evidence that both the choice of relational path and the optimization of inserted values can affect attack effectiveness.

Critical weaknesses
  1. Dependence on ground-truth labels.

The path-selection scores, RelPGD objective, and MLS all rely on the ground-truth label yiy_i. However, attackers typically DO NOT have access to test labels. Without them, it is unclear how the attacker can determine whether an attack succeeds or fails and decide when to invoke MLS. The authors should clarify how the framework operates without ground-truth labels.

  1. Sensitivity of experimental results.

The main paper reports a 10.1% (25.6-15.5) improvement in macro ASR over W-norm using a single seed, whereas the 3-seed results in Appendix Table 7 show a smaller improvement of 7.65% (23.50-15.85). This raises concerns about the stability of the reported gains. Multi-seed results should be emphasized in the main paper.

  1. Limited scalability.

The framework optimizes attacks independently for each target, requiring repeated graph reconstruction and model evaluations. This may incur substantial computational overhead when attacking large numbers of targets? The paper does not sufficiently evaluate its scalability in such scenarios.

Questions for authors
  1. How efficient is the model during an attack?

  2. How robust is the model to random seeds, and how sensitive is it to hyperparameters?

  3. How can MLS be implemented in the evaluation without knowing Y?

Reviewer_QQ8POpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks whether someone who can only add new rows to a database, say by registering an account and posting a review, can flip a relational deep learning model’s prediction about an entity they cannot touch. The authors treat this as a constrained form of graph injection against a frozen model the attacker can fully inspect: every new node has to be a legal row, every new edge a valid foreign key, and the timestamps have to be consistent. They split the attack into two steps. First they choose which tables to insert into, ranking the options with scores from trial insertions and gradients. Then they choose the new rows’ values from each column’s allowed values, using a projected gradient method (RelPGD) and, on targets it fails on, local search. On RelBench tasks with two GNN models, these insertions work very well for some task–model pairs and hardly at all for others.

Critical strengths
  1. Clear, well-motivated problem with a realistic setting: the attacker can only insert rows they’re allowed to write, with valid values, keys and timestamps.

  2. Broad evaluation and a useful takeaway: respecting database integrity alone doesn’t make RDL models robust.

Critical weaknesses
  1. The gains from the two proposed methods are uneven. The path selector beats a simple weight-norm heuristic at small budgets but not at larger ones, and its combined score isn’t clearly better than single scores in the ablation. RelPGD’s edge over CAPGD mostly disappears once search is added, and plain random search is often competitive, so it is unclear how much the methods themselves contribute.

  2. The attacker is strong: full white-box access, and free choice of event timestamps, which effectively means backdating rows. There are no gray-box, transfer, detection or defense results, so the practical risk is hard to judge.

Questions for authors
  1. Do the proposed methods still help under matched cost? Specifically, how does random search compare with RelPGD and CAPGD at the same number of model evaluations, and why does the path selector lose to the weight-norm baseline at larger budgets?

  2. In realistic systems the database usually sets timestamps when a row is created. How much does attack success drop if inserted rows must carry the current time, or if the attacker only has gray-box or transfer access to the model?

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

Related papers

Discussion (0)

Sign in to comment.