acceptodds
Withdrawn from ICLR 2027

Adaptive Energy-Guided Routing for Robust Out-of-Distribution Detection with Synthetic In-Distribution Data

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Out-of-Distribution (OOD) detection under adversarial attacks is critical for safety-sensitive AI systems. Prevailing OOD-centric methods face a fundamental limit—the OOD space is infinite—while the in-distribution (ID) manifold is compact, suggesting a complementary strategy: enriching the model's ID representation with synthetic data from diffusion models. However, the effectiveness of synthetic ID data is tied to its generation quality: high-fidelity synthetic samples effectively improve robust OOD detection, while low-quality samples can cause severe degradation when naively incorporated. To enable robust use of synthetic ID data across varying quality levels, we propose Adaptive Energy-Guided Routing (AEGR). AEGR uses a teacher classifier's LogSumExp energy as a per-sample fidelity oracle, computing a dynamic weight that controls a dual-objective routing mechanism: high-quality samples () undergo standard classification, while low-quality samples () are steered toward a margin-regularized uniform loss that flattens overconfident predictions and penalizes excessive energy. We show that the explicit margin is structurally justified by a gradient zero-sum property of the uniform loss; the reject-class paradigm avoids this requirement through its implicit energy suppression. Experiments across four datasets demonstrate that AEGR consistently improves robust OOD detection, with the largest recovery over naive synthetic data usage on the most challenging dataset where generation quality is lowest, while preserving gains on simpler datasets and relaxing the margin when generation quality is sufficient.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Robust Out-of-Distribution Detection, Adversarial Robustness, Synthetic Data

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

This paper investigates the impact of synthetic in-distribution (ID) data on adversarially robust out-of-distribution (OOD) detection, specifically addressing how the direct inclusion of low-quality synthetic data can degrade detection performance. The authors propose AEGR, a method that employs a frozen "teacher" model—pre-trained via Outlier Exposure (OE)—to compute routing weights for synthetic samples based on their Log-Sum-Exp (LSE) energy. These weights are then used to balance the classification loss against a uniform loss (incorporating an energy margin). Finally, a "student" model undergoes adversarial training using a combination of real ID data, auxiliary OOD data, and synthetic ID data. The authors validate their approach on CIFAR-10, CIFAR-100, Tiny-ImageNet-200, and ImageNet-200, providing detailed analyses on factors such as routing thresholds, energy margins, teacher configurations, data filtering, and training overhead. The most notable gains are observed on the Tiny-ImageNet-200 dataset, where AEGR effectively recovers the performance lost due to naive synthetic augmentation and outperforms baselines that do not utilize synthetic data.

Critical strengths
  • The research problem is practically significant, and the paper demonstrates targeted improvements. Synthetic data does not necessarily enhance robust OOD detection; in fact, its direct inclusion can lead to noticeable performance degradation. As shown in Table 1, the TAPGD AUC for Tiny-ImageNet-200 drops from 93.61 (with AT-OE) to 78.61 (with naive augmentation), whereas AEGR achieves 95.61. This indicates that the authors have addressed a specific failure mode, with the reported gains representing more than just a recovery to the original baseline.
  • The loss design is grounded in a clear, mechanistic rationale. Section 5.3 points out that a uniform loss is invariant to a common shift in logits, and thus fails to directly constrain the absolute LSE energy used for detection. Introducing an energy margin provides a concrete constraint to address this gap. This design can be integrated into existing training pipelines, and since inference requires only the student model, it offers clear practical utility.
Critical weaknesses
  • The core "quality-oracle" assumption lacks sufficient independent verification. The paper attributes synthetic data quality primarily to differences in "teacher energy"; however, a lower teacher score could also reflect the teacher's inherent recognition limitations or an inability to handle specific generative features. The main text and appendices fail to provide concrete examples—such as specific images, class predictions, or failure cases—associated with different routing weights, making it difficult for readers to verify this interpretation. While Appendix H.3 compares an OE-trained WRN-40-4 with a ResNet-18, this limited architectural coverage is insufficient to rule out the influence of the teacher's own capabilities on the conclusions. This issue is critical: it determines whether AEGR truly achieves reliable "quality-conditioned routing," or if it merely redistributes training objectives based on the preferences of a specific classifier.

  • Furthermore, the primary attack objective does not directly correspond to the LSE detection score actually employed. As defined in Appendix D, Eq. (11), for an OE-style detector, the attack maximizes (zt−E)(z_t - E) —where E=log⁡∑jezjE = \log \sum_j e^{z_j} —which is mathematically equivalent to maximizing log⁡pt\log p_t (the softmax probability of the target class). The paper, however, classifies higher LSE energy as "in-distribution"; these two criteria are not equivalent. Specifically, shifting all logits simultaneously can alter the LSE energy without affecting the softmax probability. While the current attack might still be effective, the lack of verification specifically targeting the LSE score means the primary results do not sufficiently substantiate the claimed robustness of the energy-based detection method.

Questions for authors
  • Is there any evidence—independent of "teacher energy"—demonstrating that synthetic samples with low scores are truly unsuitable for ID augmentation, rather than simply being valid ID samples that the teacher failed to identify? Please provide representative examples across different γ\gamma ranges, including the generated class, the teacher’s prediction, and instances where routing was incorrect or ambiguous. Also, please clarify the selection criteria for these examples. This will help determine whether the "quality-oracle" hypothesis holds.

  • Keeping the student model and synthetic dataset constant, does employing a stronger teacher significantly alter routing decisions or final performance? While the two OE-trained residual CNNs in Appendix H.3 offer preliminary evidence, incorporating a more powerful teacher—subjected to comparable OE training—would provide a more direct test of how sensitive the conclusions are to the teacher’s capability. A ViT could be one option; the primary objective is to verify the teacher’s reliability, rather than mandating a specific architecture.

  • Does the actual implementation of the TAPGD objective align precisely with Eq. (11)? Furthermore, has an attack that directly maximizes the LSE energy of OOD samples been evaluated? If not, please provide results for this approach—aligned with the actual detection score—to confirm whether the reported improvements remain robust.

  • Eq. (5) in the main text uses γ=σ(mout−ET)\gamma = \sigma(m_{\mathrm{out}} - E_T), whereas Appendix C introduces a temperature parameter τ\tau. Which formulation is used in the actual implementation, and how is τ\tau set across the various experiments? To resolve any ambiguity in the current description, please illustrate the full computational pipeline—from teacher energy and routing weight to student loss—using one of the aforementioned example samples.

Reviewer_MC3qOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the use of synthetic in-distribution (ID) data for adversarially robust out-of-distribution (OOD) detection. The authors observe that high-quality synthetic data can improve detection performance, while low-quality samples may significantly degrade robustness. To address this issue, they propose Adaptive Energy-Guided Routing (AEGR), which uses a pretrained teacher model's energy scores to assess synthetic samples and dynamically balance classification and uniform losses. An additional energy margin is introduced to suppress excessive confidence in low-quality samples. Experiments on four datasets show that AEGR improves robust OOD detection compared with directly incorporating synthetic ID data.

Critical strengths

S1. The paper presents an interesting method that is worth exploring further in future research. S2. The method seems to have a positive effect.

Critical weaknesses

W1. The method relies heavily on teacher energy scores to assess synthetic sample quality, but low energy does not necessarily indicate poor generation quality. Some valid but challenging ID samples may also receive low scores, causing them to be treated as low-quality samples and potentially limiting the model's robustness. I suggest that the authors conduct a small human evaluation by sampling synthetic images from different energy ranges and assessing their visual quality and label correctness. This would help verify whether energy is a reliable indicator of synthetic data quality.

W2. As shown in Table 1, although AEGR improves robust OOD detection, it loses most of the classification accuracy gains from synthetic data on Tiny-ImageNet-200. ID accuracy drops from 52.54% to 46.35%, nearly matching the original baseline of 46.02%. The paper should further discuss this trade-off and explore whether a better balance between ID accuracy and OOD robustness can be achieved.

W3. The choice of routing threshold is not sufficiently explained. The paper uses very high percentiles (99.5 or 99.99) of real ID energy scores to define the threshold. According to Eq. (5), this could assign large routing weights to many synthetic samples, including high-quality ones. It would be helpful to report the actual distribution of routing weights and clarify how much synthetic data contributes to each training objective.

W4. Section 5.4 suggests that the margin mainly affects samples routed to the OOD-oriented track, but Eq. (6) applies it to all synthetic samples regardless of their routing weights.

W5. Clarify the implementation of SIO and the synthetic-to-real data mixing ratio used in the experiments.

W6. Section 6.1 lists OE as a baseline, but its results are not separately reported in Table 1. The authors should clarify this omission.

Questions for authors

Q1. Could the authors report results from multiple random seeds in Table 1, including the mean and standard deviation, to demonstrate the stability of the improvements?

Q2. Could the authors improve the layout of figures and text? Some parts of the paper have formatting issues that affect readability.

Q3. Please refer to the weaknesses discussed above.

Reviewer_Hf6zOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The proposed paper studies the use of in-distribution (ID) data (generated by diffusion models) for robust Out-of-Distribution (OOD) detection under adversarial attacks. The authors observe that naive synthetic data augmentation improves robust OOD detection on CIFAR-10/100 but substantially degrades performance on Tiny-ImageNet-200 (93.61% to 78.61% TAPGD AUC). TAPGD is a multi-targeted auto-PGD, where PGD denotes Projected Gradient Descent) and AUC is the area under the Receiver Operating Characteristic (ROC) curve. To address this issue, the authors propose Adaptive Energy-Guided Routing (AEGR), which uses an Outlier Exposure (OE) teacher's LogSumExp score to compute a per-sample routing weight. This weight interpolates between cross-entropy loss and a uniform OE loss, with an additional energy margin. Experiments cover four datasets and a reject-class generalization (reported for CIFAR-100 and Tiny-ImageNet-200 only). The experiments show that AEGR improves robust OOD detection over AT-OE and AT-OE + SynthID, but the gains are small on three of the four datasets (TAPGD AUC: CIFAR-10 89.36 vs. 89.28 for SynthID, CIFAR-100 70.15 vs. 69.46, ImageNet-200 64.88 vs. 64.09 for AT-OE), while there is a large gain on Tiny-ImageNet-200 (95.61 vs. 78.61 for SynthID and 93.61 for AT-OE).

Critical strengths

Strength 1:

The observation that naive synthetic ID data can strongly degrade robust OOD detection on Tiny-ImageNet-200 (93.61 to 78.61 TAPGD AUC).

Strength 2:

A gradient analysis of the uniform Outlier Exposure (OE) loss, i.e., the zero-sum gradient property in Eqs. 9 and 10.

Critical weaknesses

Weakness 1:

I am confused about the consistency between the detection score and the adversarial evaluation. My first concern is that the paper calls its score "LSE energy" (p. 3, Sec. 3.1, Eq. 1, L. 136-144) and states that "samples with higher LSE energy are considered in-distribution." However, in Liu et al. (NeurIPS 2020), whom the authors cite, the energy is the negative LogSumExp of the logits (Sec. 2, Eq. 4), and the OOD score is the negative energy (Sec. 3.1). The paper's score is therefore identical to Liu et al.'s negative energy score (with T = 1), so the paper's "LSE energy" is the negative of the energy defined by Liu et al. The paper is internally consistent (higher LSE means ID), but using the word "energy" for the negated quantity reverses the sign convention of Liu et al. and can confuse readers. The authors should state this explicitly, for example by calling the score "negative energy" or "LSE score". Liu et al. write that "Examples with higher energies are considered as OOD inputs and vice versa" (p. 4, two lines above Eq. 6), which means that higher LSE indicates in-distribution (ID). To double check this concept, Liu et al. also wrote: “E(x; f) tends to be lower for in-distribution data” (p. 4), so lower energy indicates ID and, equivalently, higher LSE indicates ID. More importantly, since higher LSE indicates ID, an attack on this detector should increase LSE. The adversary's goal of making OOD samples appear ID is consistent with prior work (Augustin et al., 2020). However, I am confused by the statement that this is achieved by "suppressing the detection signal. (Appendix D, L. 787-788)" Since the paper explicitly defines higher positive LSE as more ID-like (Equation 1), what exactly is being suppressed? Furthermore, Eq. 11 maximizes z_t - E. If E is the LSE of Eq. 1, then z_t - E = log softmax_t. This raises the target logit, but also lowers all other logits, so it does not directly maximize LSE and does not guarantee that LSE increases. The phrases "suppressing the detection signal" (L. 787-788) and "minimizing the OOD detection score" (L. 755-756) only make sense if E is the negative-LSE energy of Liu et al. (lower means ID), which is however the opposite of Eq. 1. The authors should clarify which sign E has in Eq. (11), and whether the attack is meant to increase the LSE of the OOD input?

Weakness 2:

The main part of the proposed paper lacks an ablation that isolates per-sample adaptive routing. AEGR introduces two changes at once: (1) combining classification and OOD-oriented losses for synthetic samples, and (2) using teacher energy to assign a different mixing weight γ to each sample. While the experiments show that AEGR improves performance, they do not clearly establish which of these two changes drives the improvement. A constant gamma baseline would separate these effects: it would retain the same loss combination but assign the same gamma to every synthetic sample.

A second, related issue concerns the definition of the routing boundary m_out. The inconsistency also extends to the paper's stated motivation and contributions. In the following I use two labels to express this inconsistency. Group A: m_out is an upper-tail percentile, i.e., the p_out-th percentile of the teacher LSE on real ID data with p_out = 99.5 or 99.99, as written in Eq. 4 and Algorithm 1. Group B: m_out is a lower-tail percentile, i.e., the (100 - p_out)-th percentile. The abstract (p. 1, L. 21-25) and Introduction (p. 2, L. 65-76) emphasize that high-quality synthetic samples should receive gamma ≈ 0 and undergo standard classification, while low-quality samples should receive gamma ≈ 1 and undergo OOD-oriented regularization. The Contributions section (p. 2, L. 81-94) describes the same mechanism. Since γ = σ(m_out − E_T) in Eq. 5, this description holds for any value of m_out and therefore does not fix the tail. It is only meaningful if m_out is a lower-tail percentile (Group B). The supplemental code confirms this: in train_unif.py (L. 451-454) m_out is the (100 − p_out)/100 quantile of the teacher LSE on real ID data minus ε = 1, i.e., a lower-tail percentile. This is also how L. 243-244, L. 428 and Figure 1(b) read. In Figure 1(b), the boundary m_out (black dashed line) lies below the real ID mean energy of 14.44 (L. 203), which is inconsistent with an upper-tail percentile at the stated p_out = 99.5 (Group A) and consistent with a lower-tail percentile (Group B). In contrast, Eq. 4 and Algorithm 1 define m_out as the p_out-th percentile, which with the stated defaults (99.5 and 99.99) reads as an upper-tail percentile (Group A). Thus, the introduction, contributions, equations, algorithm, and implementation do not consistently describe the same routing mechanism. Although the code strongly suggests that Group B is used to produce the results, the authors should explicitly clarify these conflicting descriptions.

Questions for authors
  1. Is m_out the upper-tail percentile (as in Eq. 4 and Algorithm 1) or the lower-tail percentile (as in Figure 1(b), L. 243-244, L. 428 and the released code, which also subtracts ε = 1)? Could the authors correct Eq. 4, Algorithm 1, L. 438 and Appendix B accordingly, and report the numerical values of m_out per dataset for the default p_out?

  2. Could the authors compare AEGR with constant γ settings (e.g., γ = 0.88, which is the mean routing weight in Appendix C, as well as γ = 0.5 and γ = 1), on all four datasets?

  3. Could the authors explicitly clarify that their positive LSE detection score is the negative of the conventional energy defined by Liu et al. (NeurIPS 2020)?

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

Related papers

Discussion (0)

Sign in to comment.