acceptodds
Withdrawn from ICLR 2027

The Imperfect Mirror: Non-intrusive Model Auditing via Information Bottleneck Deterioration

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing non-intrusive model auditing methods mainly rely on fingerprinting, which identifies stolen models through behavioral similarity preserved from the victim model. However, this similarity is fragile under practical stealing and evasion, such as cross-architecture extraction and adversarial training. Given these limitations, we propose a novel non-intrusive framework that shifts auditing from preserved similarity to structured information deterioration. Our key insight is that independently trained models tend to exhibit a characteristic information-bottleneck balance, reflecting how they compress input information while preserving label-relevant information. In contrast, model stealing can disrupt this learned balance, leading the stolen model to retain more input information while preserving less label-relevant information than the victim, which provides a principled auditing signal. We formalize this insight into a two-stage auditing procedure that first screens for this directional deterioration and then statistically assesses whether the suspect's information-bottleneck behavior is consistent with that of independently trained models, yielding a quantitative auditing score. Experiments demonstrate effective auditing with low false-positive rates and improved robustness over the evaluated baselines under cross-architecture stealing and adversarial-training evasion.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Model Auditing, Information Bottleneck

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_5Bt5OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper considers a model auditing scenario where a model owner has a victim model and suspects that another model may be derived from it. The auditor only has black-box access to the suspect model and wants to determine whether it was stolen without modifying the original victim model.

The paper addresses this problem by looking at how the suspect model differs from the victim. The main observation is that stolen models tend to retain more input information but less label-related information than the victim. Based on this, the method first checks whether the suspect shows this directional information deterioration, and then compares it with a group of independently trained models to determine whether such behavior is unusual for a normally trained model.

The experiments cover multiple model architectures and stealing strategies, including fine-tuning, pruning, knowledge distillation, and model extraction. The proposed method achieves strong detection performance.

Critical strengths
  1. The information deterioration phenomenon identified in this paper is interesting and insightful. Specifically, the authors observe that stolen models tend to retain more input information while preserving less label-relevant information than the victim model. This phenomenon provides a novel perspective for model auditing.

  2. The proposed auditing method performs well across a wide range of model-stealing strategies. This consistent performance suggests that the proposed auditing signal has good generality across different stealing mechanisms.

Critical weaknesses
  1. My main concern with this paper is that the proposed method requires querying the suspect model with the victim model’s training data. In practice, the owner of the suspect model is usually an untrusted third party, and sending private training samples to such a model could expose these data to the third party. I believe this introduces an additional privacy and security risk. It seems unreasonable to require the model owner to expose private training data in order to verify model ownership.

  2. It is unclear how universal the observed information deterioration phenomenon is. Although the paper provides some theoretical analysis, it depends on several specific assumptions that may not hold for model stealing in general.

  3. The evaluated model architectures are somewhat outdated. Most experiments are conducted on relatively small or classical vision models. It would strengthen the paper to evaluate whether the proposed information-bottleneck deterioration phenomenon also holds for more recent and larger vision architectures.

Questions for authors
  1. Can the proposed method still achieve strong auditing performance without using the victim model’s private training data? If the method can be shown to work well with non-private or held-out data, this would substantially address my main concern, and I am willing to increase my score.

  2. Can the information deterioration phenomenon also be observed in more modern architectures?

Reviewer_Si1zOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a black-box method to audit whether a suspect classifier was derived from a victim through model stealing (e.g., fine-tuning, pruning, distillation, or extraction). On the victim’s private training set, it computes two information quantities from binned output probabilities, I(X;T)I(X;T) and I(T;Y)I(T;Y). The paper performs model auditing in two steps: (1) check if the suspect has higher I(X;T)I(X;T) and lower I(T;Y)I(T;Y) than the victim, and (2) a Hotelling’s T2T^2 test finds its two values atypical of independently trained reference models. Experiments on CIFAR-10/100 and ImageNet-100 report 100% detection and no false positives, including under cross-architecture stealing and adversarial training.

Critical strengths
  1. The formulation of model auditing from the information bottleneck perspective is quite interesting. The motivation on the cases where fingerprinting fails is valid and validated in the experiments.

  2. The auditing method is black-box and can potentially be used in practice.

  3. I briefly checked the proofs of the theoretical results, and they look correct and sound.

  4. The two-stage scoring method is interesting and has a valid p-value under the stated null assumptions.

  5. The experiments are extensive overall.

Critical weaknesses
  1. Figure 1 is very hard to understand without reading the entire paper. For example, I,X,T,YI, X, T, Y are not defined in Figure 1, and different fonts are used in the left figure and in other parts of the paper. Also, TT denotes Hotelling's T2T^2 test, and I would suggest using a different notation for the model's representation.

  2. Similarly, there are several undefined terms in Section 1, including “learned information processing,” “information-bottleneck balance,” and “structured information deterioration”, which make it hard to read.

  3. The definition of the representation TT in Section 4.1 (T=f(X;θ)T = f(X; θ), which seems to be a feature vector) does not match what the method actually computes. After digging through the appendix, it is clear that TT is the output probability vector discretized by a binning map QQ, but QQ is introduced only in Appendices A.2 and B.1. I believe the definition of TT is very important, as it determines both I(X;T)I(X;T) and I(T;Y)I(T;Y) and should be stated in the main text.

  4. Theorems 4.2 and 4.3 depend on Assumptions A.1–A.3, but it seems that the paper does not verify them properly. A.1 seems to depend on the victim memorizing DvD_v; A.2 is never verified but seems to be computable; and A.3 is violated at PR-80% pruning (Figure 4). Thus, I find the theoretical support for the deterioration claim rather weak.

  5. The low false-positive claim is tested only on negative suspects generated by the same data, architecture, and training protocol as the reference pool. Thus, the experimental results seem to be expected, as the Stage 2 null holds by the construction of the experiments. In realistic settings, innocent models can be trained differently (e.g., more data, pretraining, different recipes, an independent teacher, or partial overlap with DvD_v).

Questions for authors

Please see the weakness above. I expect the authors to improve the presentation of the paper and directly address W4 and W5.

Moreover, the proposed method requires training independent reference models. Does this introduce significant additional computational overhead?

Reviewer_fmT9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how to detect model stealing without modifying the protected model and with only black-box query access to the suspect model. The authors propose a model auditing framework based on information-bottleneck deterioration, which identifies characteristic information changes potentially induced by model stealing. Experiments are conducted on CIFAR and ImageNet-100. The results show that the proposed method achieves higher detection rates than existing model fingerprinting methods, particularly under cross-architecture model stealing and adversarial-training-based evasion.

Critical strengths

The paper focuses on changes in information processing induced by model stealing, rather than relying on preserved behavioral similarity between the victim and suspect models. This approach provides an alternative to conventional fingerprint matching, providing meaningful methodological novelty.

Critical weaknesses
  1. The proposed method relies on information-bottleneck deterioration as a key indicator of model stealing. However, similar information-bottleneck deterioration could also arise from general model performance degradation, differences in training strategies, or changes in output distributions, rather than necessarily indicating that a model has reused the victim's parameters or predictions. Therefore, it remains unclear whether the proposed information-bottleneck deterioration is specifically associated with model stealing or merely reflects more general differences in model behavior.

  2. The adversarial robustness evaluated in the paper are primarily based on adversarial training, rather than adaptive attacks specifically designed to circumvent the proposed information-bottleneck auditing mechanism. An attacker could potentially manipulate those information-theoretic statistics through probability calibration or other output post-processing techniques. The paper does not evaluate such auditing-aware attack strategies, leaving the robustness of the proposed method against adaptive attackers in question.

Questions for authors

I am wondering whether information-bottleneck deterioration can also result from general model performance degradation, changes in output distributions, or differences in training conditions, rather than model stealing. Additional controlled experiments are needed to distinguish stealing-specific information deterioration from more general variations in model behavior.

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

Related papers

Discussion (0)

Sign in to comment.