acceptodds
Under review as a conference paper at ICLR 2027

VeriAudit: An Adversarial Neuro-Symbolic Framework for Industrial Data Product Auditing

Abstract

As the digital economy shifts from raw data exchange to the circulation of Derived Data Products (DDPs), a core question arises: how can providers trade insights without implicitly leaking proprietary assets? While Trusted Execution Environments (TEEs) provide a secure hardware container for computation, they often suffer from a Semantic Gap, certifying code integrity but remaining blind to high-order statistical leakage, which limits the effectiveness of traditional post-hoc sampling against sophisticated inference attacks. To address this, we propose VeriAudit, an adversarial neuro-symbolic infrastructure integrating Continuous Auditing, Full-population Auditing, and AI Auditing within industrial data ecosystems. A "Dynamic Policy Injection" architecture enables "hot-patching" of audit logic inside running TEEs via remote attestation, without service interruption. Its core combines symbolic constraints for deterministic compliance with a lightweight GNN for adaptive pattern recognition, optimized for TEE memory limits. Evaluations on an industrial Energy-Twin dataset show that VeriAudit achieves a 99.0% recall rate against adaptive attacks where traditional rules exhibit limitations. Sustaining 10,000 TPS at 20ms latency, VeriAudit serves as a practical paradigm for audit-aware data infrastructure in the modern data factor market.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Trusted Execution Environments, neuro-symbolic, data auditing, derived data products, statistical leakage

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_gEDiOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces VeriAudit, a system designed to detect "semantic leakage" prior to the release of data products. The system combines fixed, rule-based checks—which identify clear numerical anomalies—with a three-layer Graph Convolutional Network (GCN) that uncovers suspicious correlations between raw data and publicly disclosed metrics. Additionally, the authors propose a policy update strategy and a "Proof-of-Audit" mechanism tailored for Trusted Execution Environments (TEEs).

Experiments were conducted using synthetic electricity consumption data from ten factories. Public metrics were generated via a weighted moving average, with simulated attacks inducing leakage by manipulating the aggregation weights. The authors report a recall of 99.0%, with performance benchmarked primarily against rule-based methods. The TEE environment was implemented via software simulation.

Critical strengths

Conducting a joint audit prior to data release is supported by sound system design rationale. By examining both the raw data and the intended output, auditors can identify correlations that would be difficult to discern by looking solely at the public-facing results. Furthermore, integrating rule checks, model verification, and audit logging into a unified workflow provides a clear, prototypical framework.

Critical weaknesses
  • The experimental scenario is overly simplistic, and the relationship between the attack setup and the threat model remains unclear. The experiments consider only a basic aggregation of electricity consumption data from ten factories. The attack involves modifying the aggregation weights so that the output becomes highly correlated with the data of a specific factory. This is insufficient to represent the broader range of industrial data auditing issues discussed in the paper. Furthermore, while §3.2 and Appendix B primarily describe a scenario where a malicious buyer infers secrets through repeated queries, the attacker in the experiments is capable of modifying internal aggregation weights. Additionally, Figure 3 labels the data-processing entity as a "Malicious Engine." The paper fails to clearly define the relationships between these roles or explain how a buyer could obtain the authorization to modify aggregation weights. Consequently, it is difficult to determine whether the experiments actually correspond to the proposed threat model.

  • The baselines provided are inadequate to demonstrate the necessity or the added value of the GCN-based approach. The attack described in Appendix D exhibits a distinct characteristic: the Pearson correlation between the output and the target factory's data exceeds 0.95. Therefore, a sliding-window correlation-based detection method would serve as a direct and significant baseline; however, the paper does not include such a comparison.The primary baseline only checks for numerical ranges and the magnitude of fluctuations. Since the authors explicitly state that the attack is designed to bypass these checks, it is not surprising that fixed-rule-based methods yield a recall of 0%. The current comparisons fail to illustrate the benefits of the GCN approach over simple statistical methods, nor do they sufficiently substantiate the paper's core methodological contributions.

  • Detecting anomalous correlations does not, in itself, prove that information leakage has been prevented. The paper does not fully demonstrate how an attacker recovers secrets from the published results, nor does it report the accuracy of such recovery before and after the defense is applied. The current "recall" metric measures the ability to identify injected attacks, but it does not directly quantify the extent to which the system prevents actual information leakage.Moreover, the paper links the GCN risk score to Mutual Information without providing an adequate theoretical basis. This creates a significant gap between the privacy objectives defined in the problem formulation and the actual evaluation performed.

Questions for authors
  1. Is the attacker merely a data buyer who can only query the results, or a data processor capable of controlling the computation process? Where does the authority to modify aggregation weights in the experiments originate? Please clarify how the experimental attacks map to the threat model described in §3.2 and Appendix B.

  2. Have you compared your approach against simple statistical methods, such as a sliding window Pearson correlation? What performance gains does the GCN offer at a comparable false positive rate? How do the current experiments demonstrate that the model identifies leaks beyond those characterized by obvious, manually defined strong correlations?

  3. Have you implemented an actual attack to recover private data from the publicly released results? How do the attack success rate or the reconstruction error change when VeriAudit is enabled versus disabled? Furthermore, is the GCN risk score an empirical risk metric, or does it provide a formal guarantee regarding the mutual information condition specified in §3.2? If the latter, please provide the supporting rationale.

Reviewer_WCtLOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes VeriAudit, a framework for detecting semantic information leakage in derived industrial data products. It combines rule-based checks with a lightweight graph neural network inside a trusted execution environment to identify suspicious statistical dependencies between sensitive inputs and released outputs. The framework also supports dynamic policy updates and verifiable audit records. Experiments on a synthetic industrial energy dataset report 99.0% recall, 96.2% precision, 20 ms latency, and 10,000 transactions per second, along with analyses of computational efficiency and resource constraints.

Critical strengths
  1. Clear problem motivation. The paper highlights that protecting the integrity and confidentiality of computation does not necessarily prevent sensitive information from being inferred from released data products. Extending security auditing to the semantic properties of outputs is a relevant direction for trustworthy industrial data sharing.
  2. Practical system-level considerations. The framework attempts to jointly address leakage detection, real-time auditing, resource constraints, and audit accountability. The inclusion of latency measurements and resource sensitivity analyses is a positive effort toward evaluating practical feasibility.
Critical weaknesses
  1. Mismatch between the threat model and the evaluated attacks. The paper describes a malicious consumer who tries to infer sensitive information by repeatedly querying the system. However, in the experiments, attacks are generated by manipulating aggregation weights to make the public outputs strongly correlated with private inputs. I am not sure how this attack fits the stated threat model, since these weights would normally be controlled by the data processor rather than the consumer. Without a clearer explanation of the attacker's capabilities, it is difficult to tell whether the evaluation reflects the intended security setting.
  2. Insufficient justification for using GCN. The main detection task seems to involve identifying unusually strong correlations between private inputs and public outputs. Such patterns might also be detected by much simpler methods, such as sliding-window Pearson correlation analysis. However, the paper mainly compares VeriAudit against static numerical rules, which are not designed to capture these correlations. I would like to see comparisons with stronger but still lightweight baselines to understand what the GCN actually contributes.
  3. Unclear connection between anomaly detection and privacy protection. The experiments mainly evaluate how accurately VeriAudit identifies injected attacks, using recall and precision. However, detecting suspicious correlations does not necessarily mean that sensitive information can no longer be inferred from the released outputs. The paper does not evaluate whether VeriAudit actually makes private data reconstruction more difficult for an attacker. Also, while the problem formulation introduces a mutual information constraint, it is unclear how the learned risk score relates to this objective. As a result, the reported detection performance alone does not fully support the claimed privacy protection.
  4. Incomplete experimental validation and questionable efficiency claims. The paper studies the effects of GCN depth and historical window size, but provides limited ablations of the main design choices, making it difficult to assess which components are actually responsible for the reported improvements. I also noticed a potential inconsistency in the efficiency results. The paper reports 10,000 TPS, while Appendix C states that GCN inference takes 9 ms per batch of 64 transactions and is executed serially. This would limit the throughput of that stage to roughly 7,111 TPS. It is unclear how the reported throughput is achieved under these constraints.
Questions for authors
  1. Threat model: Could the authors clarify how the malicious consumer is able to manipulate aggregation weights in the evaluated attacks? Is this allowed through the query interface, or does it require additional privileges not specified in the threat model?
  2. Choice of GCN: Have the authors considered simpler baselines, such as sliding-window correlation analysis or lightweight temporal anomaly detectors? What advantages does graph-based modeling provide over these methods in the current setting?
  3. Privacy protection: Can the authors provide evidence that VeriAudit actually reduces an attacker's ability to reconstruct sensitive data? Also, how should we interpret the learned risk score in relation to the mutual information constraint in Section 3.2?
  4. Experimental validation: Could the authors clarify the contribution of the main components through additional ablations? In particular, how is the reported 10,000 TPS achieved given the serialized GCN inference described in Appendix C?
Reviewer_GBEwOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes VeriAudit, a neuro-symbolic auditing framework that combines symbolic rules with a lightweight GNN to detect semantic leakage in derived data products.

Critical strengths
  • Interesting topic
  • Reasonable hybrid architecture
Critical weaknesses
  • Poor presentation
  • Weak experimental settings
Questions for authors

This paper proposes VeriAudit, a neuro-symbolic auditing framework that combines symbolic rules with a lightweight GNN to detect semantic leakage in derived data products. This is an interesting topic; however, the paper is not complete, and the presentation is poor.

There are several main concerns.

  • The experimental settings are too weak. There is only one synthetic dataset and only one proposed attack without baselines.
  • The TEE is simulated. There is a lack of practical validation.
  • Many numbers and settings are not consistent.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.