acceptodds
Under review as a conference paper at ICLR 2027

FedOPhD: Federated Post-Hoc OOD Detection via Unlearning-Inspired Probing

Abstract

In federated learning (FL), in-distribution (ID) data include both Local-ID (L-ID) from the deployment client and Transferred-ID (T-ID) from other participating clients. Locally unfamiliar T-ID must be distinguished from true out-of-distribution (OOD) inputs, even when only a trained global model and L-ID data are available. We propose FedOPhD, a post-hoc federated OOD detector that probes semantic support through unlearning responses. FedOPhD performs client-conditioned semantic attenuation on a diagnostic model copy, using the client's learned label support to attenuate locally supported semantics. It measures how an input's representation changes align with responses on an L-ID reference subset across multiple layers. These responses offer additional evidence for distinguishing inputs with similar original model prediction confidence. This response evidence is combined with the original model's global recognition support through locally calibrated fusion, accounting for T-ID beyond local semantic coverage. Detector construction requires neither OOD data nor further federated communication, while the deployed classifier remains unchanged. Experiments on CIFAR-10 and CIFAR-100 demonstrate the effectiveness of FedOPhD in distinguishing OOD inputs from both L-ID and T-ID across heterogeneous federated settings.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Post-hoc OOD detection, Federated unlearning, Federated OOD detection

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_gQeqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies post-hoc out-of-distribution detection after federated training has ended. A deployment client has access only to the trained global classifier and its own in-distribution data, but should also accept legitimate ID inputs originating from other clients, including classes absent from its local data. FedOPhD constructs a diagnostic copy of the model through a short sequence of client-conditioned semantic-attenuation updates. It measures representation changes across intermediate layers, fits local geometric response modes, and uses prediction-margin changes to construct functional routing directions. The resulting response-alignment support is combined with energy-based recognition support from the unchanged global classifier through a locally calibrated, response-dependent fusion rule. The evaluation covers near- and far-OOD detection on CIFAR-10/100, transferred-ID preservation, DomainNet10 under FedAvg and FedBN, additional backbones and federation configurations, and extensive intervention, readout, and fusion controls.

Critical strengths
  1. The deployment problem is well motivated and precisely formulated.

  2. The principal empirical gains are substantial under shared-checkpoint comparisons.

  3. The intervention and readout controls provide meaningful evidence for the proposed mechanism.

Critical weaknesses

W1. Federation-level ID preservation remains conditional on local response coverage, and this boundary is not fully characterized under domain shift: The fusion rule gives global recognition a response-dependent role: sufficiently weak response support can force rejection regardless of the global-recognition score. Appendix A.4 explicitly establishes this property in Proposition A.3/Eq. (23). Consequently, a globally recognizable T-ID input whose response is weakly coupled to the local intervention, or poorly represented by the locally fitted modes, may still be rejected. Proposition A.4 appropriately makes transfer conditional on response coverage and route agreement, but does not establish how frequently these conditions hold in deployment. The existing T-ID and class-coverage experiments in Tables 6–10 provide useful evidence. However, the DomainNet10 experiments in Tables 25–27 use FPR95 and T-FPR95 thresholds selected to detect at least 95% of the evaluation OOD population. They do not extend the Table 6 analysis to cross-domain T-ID at the threshold that a receiving client can actually calibrate using only local ID. Thus, the evidence for cross-domain ranking improvements is stronger than the evidence for the preservation–detection trade-off at the deployed local threshold. A targeted evaluation under both FedAvg and receiving-client FedBN, together with an analysis of T-ID samples satisfying the response-only rejection condition, would clarify this important methodological boundary.

W2. The fusion comparisons do not fully isolate query-dependent weighting from a favorable average weighting of the two signals: Tables 12–13 show consistent improvements over equal-weight arithmetic and geometric averaging and the other tested combining rules. However, the equal-weight comparisons change both the query dependence of the gate and the overall allocation of weight between response and recognition support. This distinction matters because the two channels have substantially different near- and far-OOD behavior. For example, on CIFAR-100 with α=0.1\alpha=0.1, near-OOD AUROC is 58.15 for FedOPhD-R, 72.05 for FedEnergy, and 76.79 for the full method. A more targeted control would freeze the arithmetic gate in Eq. (5) to its mean value over the same client-local reference set. Specifically, with wk(x)=sR(x)/(sR(x)+βk)w_k(x)=s_R(x)/(s_R(x)+\beta_k), one could compare against

\frac{1}{|D_{\mathrm{ref},k}|} \sum_{x\in D_{\mathrm{ref},k}}w_k(x), \qquad B_{\mathrm{fixed}}(x)= \bar w_k s_G(x)+(1-\bar w_k)s_R(x).$$ This control requires no OOD-based tuning and preserves the same fitted component scores and average reference-set weighting. It would make the attribution of gains specifically to query-dependent fusion more precise. W3. Selected fusion-baseline and sparse-client implementation details remain underspecified: Despite the extensive appendix, the GLRT and BH fusion baselines are presented primarily through their names and numerical results. The exact input transformations, combining statistics, parameter choices, and calibration-data roles should be specified. For the BH comparator in particular, the paper should explain how the component scores are converted into the conformal p-values used by the procedure; the response support itself is an aggregate of empirical ranks. The sparse-client handling also needs a more explicit algorithmic specification. Appendix D.3 states that the nominal number of response modes is reduced to the “largest locally supported” count, but the support criterion and complete rule for choosing that count are unclear. This choice affects the large-federation and reduced-local-data experiments. Providing these details would improve the reproducibility and interpretation of the relevant comparisons.
Questions for authors
  1. Local-threshold behavior under domain shift: Can you apply the Table 6 deployment protocol to DomainNet10 under both FedAvg and receiving-client FedBN, reporting L-ID false rejection, cross-domain T-ID false rejection, and OOD detection at thresholds calibrated only from receiving-client ID data? A breakdown by receiving domain would be especially informative. Among rejected T-ID samples, how many satisfy the forced-rejection condition in Eq. (23), particularly when the original classifier predicts them correctly with strong global-recognition support?

  2. Attribution to adaptive fusion: How does the proposed fusion compare with a query-independent gate matched to its mean local-reference weight, as defined above? Reporting the comparison across the same seeds and both OOD regimes would directly assess the contribution of query dependence.

  3. Reproducibility of selected comparisons: Please provide the precise GLRT and BH implementation formulas and calibration protocol, together with the executable rule for selecting the effective number of response modes for sparse clients.

  4. Routing stability as a secondary methodological clarification: The method routes with functional directions bℓ,jb_{\ell,j} but scores against geometric directions μℓ,j\mu_{\ell,j}. A small response change near a functional-routing boundary can therefore switch between different geometric scores. The finite-gain analysis in Proposition A.2 is correctly restricted to a fixed route, and Table 14 supports the average detection performance of the two-stage design. Could you additionally report routing margins or a small diagnostic of route and score stability under minor response perturbations? This would clarify the practical relevance of the fixed-route qualification.

9 comments
AuthorsOpenReview ↗

Response to Reviewer gQeq (Part 1)

W1: Federation-level ID preservation remains conditional on local response coverage, and this boundary is not fully characterized under domain shift: ... A targeted evaluation under both FedAvg and receiving-client FedBN, together with an analysis of T-ID samples satisfying the response-only rejection condition, would clarify this important methodological boundary.

Q1: Local-threshold behavior under domain shift: Can you apply the Table 6 deployment protocol to DomainNet10 under both FedAvg and receiving-client FedBN, reporting L-ID false rejection, cross-domain T-ID false rejection, and OOD detection at thresholds calibrated only from receiving-client ID data? ... Among rejected T-ID samples, how many satisfy the forced-rejection condition in Eq. (23), particularly when the original classifier predicts them correctly with strong global-recognition support?

Response to W1 & Q1

Part 1/2: Cross-domain T-ID preservation under local thresholds.

We thank the reviewer for this thoughtful suggestion. We agree that evaluating cross-domain T-ID preservation under locally calibrated deployment thresholds provides important evidence beyond the ranking-based results in Tables 32–34.

Local-threshold evaluation. Following the Table 6 deployment protocol, we evaluate DomainNet10 under both FedAvg and receiving-client FedBN, retaining the original checkpoints, client partitions, detector configurations, and five random seeds. Each receiving client calibrates its threshold using only local ID data, targeting 95% L-ID acceptance. We then measure L-ID false rejection, cross-domain T-ID false rejection, and OOD detection across the same eight external OOD sources as Tables 32–34, without retraining or OOD-based recalibration.

Table 35: DomainNet10 deployment results with L-ID-calibrated thresholds (%, five-seed means).

AggregationMethodL-ID FRR ↓T-ID FRR ↓OOD-DR ↑
FedAvgFedEnergy4.406.8333.66
FedAvgFedViM4.9635.5346.10
FedAvgFedNCI4.775.3529.54
FedAvgFedGradPCA4.7211.6629.87
FedAvgFedRAS4.747.0736.40
FedAvgFedOPhD (Ours)4.596.0952.81
FedBNFedEnergy4.7413.6323.64
FedBNFedViM4.0242.2055.23
FedBNFedNCI4.7821.6938.99
FedBNFedGradPCA4.0510.8712.33
FedBNFedRAS4.7828.4138.98
FedBNFedOPhD (Ours)4.1910.9954.68

Notes. T-ID FRR denotes cross-domain T-ID false-rejection rate. OOD-DR denotes the OOD detection rate averaged across eight external OOD sources.

The results demonstrate a favorable preservation–detection trade-off under both aggregation protocols. Compared with FedEnergy, FedOPhD improves OOD detection by 19.15 and 31.04 percentage points under FedAvg and FedBN, respectively, while achieving lower cross-domain T-ID false-rejection rates (6.09% vs. 6.83% and 10.99% vs. 13.63%). Notably, under FedBN, FedOPhD achieves nearly the same OOD detection rate as FedViM (54.68% vs. 55.23%) while substantially reducing legitimate cross-domain T-ID rejection (10.99% vs. 42.20%). Moreover, among cross-domain T-ID samples correctly classified by the original model, FedOPhD's false-rejection rates decrease further to 2.81% and 6.13% under FedAvg and FedBN, respectively.

AuthorsOpenReview ↗

Response to Reviewer gQeq (Part 2)

W1 / Q1 (continued): Analysis of the response-only rejection condition under domain shift, including T-ID samples correctly classified with strong global-recognition support.

Response to W1 & Q1

Part 2/2: Analysis of the sufficient rejection condition.

We first clarify that our theoretical analysis does not claim unconditional T-ID preservation. Proposition A.4 explicitly characterizes transfer under response coverage and route agreement, while Proposition A.3 identifies a sufficient rejection condition when Response Alignment Support is weak. We therefore extend our experiments to quantify both the practical preservation–detection trade-off and the occurrence of this theoretically identified failure condition under domain shift.

Analysis of the sufficient rejection condition. To examine the boundary highlighted by the reviewer, we evaluate Eq. (23) at each client's locally calibrated deployment threshold:

sR,k(x)<−τkβk1+βk+τk,0<−τk<1.s_{R,k}(x) < \frac{-\tau_k\beta_k}{1+\beta_k+\tau_k}, \qquad 0<-\tau_k<1.

where βk\beta_k denotes the client-specific fusion scale and τk\tau_k the deployment threshold. We measure how frequently rejected cross-domain T-ID samples satisfy this sufficient condition, including those correctly classified by the original model with strong Global Recognition Support.

Table 36: Percentage of rejected cross-domain T-ID samples satisfying Eq. (23).

Rejected T-ID subsetFedAvg (%)FedBN (%)
All rejected cross-domain T-ID10.3617.26
Correctly classified, sG>0.75s_G>0.7567.5078.70
Correctly classified, sG>0.90s_G>0.9083.2090.50

Notes. Each percentage is conditional on the rejected T-ID subset specified in the corresponding row. sGs_G denotes locally rank-calibrated Global Recognition Support, rather than a class-posterior probability.

These results confirm the failure mechanism identified in Proposition A.3 while clarifying its practical prevalence. Among rejected T-ID samples correctly classified with strong global support, the majority satisfy Eq. (23). This concentration is consistent with the response-gated fusion mechanism: as sGs_G approaches one, the sufficient condition in Eq. (23) approaches the exact rejection boundary.

Importantly, these conditional proportions should not be interpreted as false-rejection rates for the corresponding T-ID populations. Across all rejected cross-domain T-ID samples, only 10.36% and 17.26% satisfy this sufficient condition under FedAvg and FedBN, respectively. The results therefore distinguish the existence of response-induced forced rejection from its frequency among the observed rejection events, without implying that the condition captures every possible response-related failure.

Overall, these additional experiments show that the conditional theoretical boundary does not preclude effective cross-domain T-ID preservation in practice. At locally calibrated deployment thresholds, FedOPhD achieves substantially stronger OOD detection than FedEnergy while maintaining comparable or lower T-ID rejection, even under receiving-client FedBN. The analysis also identifies where the response gate can override strong global recognition, providing a more precise characterization of the method's preservation–detection trade-off under domain shift.

We have incorporated both the local-threshold evaluation and the sufficient-rejection analysis into Appendix E.3.1 (Tables 35–36) of the updated Rebuttal Version of our manuscript.

AuthorsOpenReview ↗

Response to Reviewer gQeq (Part 3)

W2: The fusion comparisons do not fully isolate query-dependent weighting from a favorable average weighting of the two signals: ... This control requires no OOD-based tuning and preserves the same fitted component scores and average reference-set weighting. It would make the attribution of gains specifically to query-dependent fusion more precise.

Q2: Attribution to adaptive fusion: How does the proposed fusion compare with a query-independent gate matched to its mean local-reference weight, as defined above? Reporting the comparison across the same seeds and both OOD regimes would directly assess the contribution of query dependence.

Response to W2 & Q2

We thank the reviewer for this valuable suggestion. We agree that the original equal-weight comparisons do not fully isolate the contribution of query dependence from the average allocation between Response Alignment Support and Global Recognition Support. The suggested mean-matched control provides a more direct assessment of this distinction.

Mean-matched Fixed Gate. We implement the proposed control using the same model checkpoints, support scores, reference data, fusion parameters, random seeds, and evaluation protocol. Specifically, for each client k, we compute the average adaptive weight over its local reference set:

wˉk=1∣Dref,k∣∑x∈Dref,ksR,k(x)sR,k(x)+βk.\bar{w}_k=\frac{1}{|D_{\mathrm{ref},k}|}\sum_{x\in D_{\mathrm{ref},k}}\frac{s_{R,k}(x)}{s_{R,k}(x)+\beta_k}.

The fixed-gate support is then

Bfixed,k(x)=wˉksG,k(x)+(1−wˉk)sR,k(x).B_{\mathrm{fixed},k}(x)=\bar{w}_k s_{G,k}(x)+(1-\bar{w}_k)s_{R,k}(x).

This control preserves the same two support signals and their average reference-set weighting while removing only the query dependence. All weights are determined exclusively from client-local ID data, without OOD-based tuning.

Table 15: Query-dependent Adaptive Fusion vs. Mean-matched Fixed Gate (%, five-seed means).

DatasetαFusion StrategyNear AUROC ↑Near FPR95 ↓Far AUROC ↑Far FPR95 ↓
CIFAR-100.1Adaptive Fusion (Ours)86.0343.9697.1313.96
Mean-matched Fixed Gate85.3145.7792.2733.85
CIFAR-100.5Adaptive Fusion (Ours)89.1539.3098.327.66
Mean-matched Fixed Gate88.6039.6395.9817.73
CIFAR-1010Adaptive Fusion (Ours)89.5239.5798.436.72
Mean-matched Fixed Gate87.9845.6195.8425.27
CIFAR-1000.1Adaptive Fusion (Ours)76.7967.4594.0623.80
Mean-matched Fixed Gate75.0369.1187.6952.66
CIFAR-1000.5Adaptive Fusion (Ours)78.7261.7294.2921.87
Mean-matched Fixed Gate76.2864.3889.9045.60
CIFAR-10010Adaptive Fusion (Ours)79.5858.7193.5223.63
Mean-matched Fixed Gate78.0959.8889.5042.29

Adaptive Fusion consistently outperforms the mean-matched control across all six settings and both OOD regimes, improving all 24 reported AUROC and FPR95 metrics. The gains are particularly pronounced for Far-OOD, where FPR95 decreases by 10.07–28.86 percentage points, alongside consistent improvements in AUROC. Near-OOD performance also improves across all settings, although the gains are generally smaller.

These improvements are consistent with the intended mechanism of our response-dependent fusion. As defined in Eq. (5), weak Response Alignment Support limits the influence of Global Recognition Support, preventing potentially misleading global recognition from dominating the OOD score. Conversely, stronger response alignment allows global recognition to contribute more substantially. A fixed gate cannot make this distinction across queries, even when its average weight is matched to the adaptive gate. This mechanism also helps explain the larger Far-OOD improvements, consistent with the stronger Far-OOD discrimination observed in our response-only ablation.

The mean-matched comparison therefore isolates the benefit of query-dependent weighting from the average allocation of the two support signals. The consistent improvements provide direct empirical support for Adaptive Fusion as a substantive component of FedOPhD, rather than merely a favorable choice of fixed fusion weights.

We have incorporated this controlled comparison and its implementation details into Appendix E.2.2 (Table 15) of the updated Rebuttal Version of our manuscript.

Reviewer_DGfhOpenReview ↗
Rating
2
Confidence
3
Informativeness
3(+1)
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

The paper tries to improve the OOD Detection capacity of the client with the global model, local model, and local client ID.

Critical strengths

The performance achieved good performance under far OOD, but was low in near-OOD.

Critical weaknesses

In general, the paper proposed ``client-conditioned support attention” to reduce the prediction probabilities of the local available ID and in the second step they propose a way to leverage the difference between the modified model θT\theta_{T} and θ0\theta_{0} to predict the OOD. The novelty of the proposed methods is not sufficient; in simple terms, the authors consider perturbing the global model for OOD detection, which has been done in various works. The difference lies in the federated setting, and this makes the contribution of the paper incremental. In addition, the presented idea is likely to be a hand-engineering problem, rather than a research problem.

Finally, many of the authors' sentences make the paper feel awkward and unnatural. For example: "Federated learning (FL) decouples what a client has seen from what its model has learned" --> I know you are trying to present the OOD problem, but writing this makes it difficult to understand.

Questions for authors

Why is the proposed client-conditioned support attenuation necessary? The authors design a target distribution that suppresses locally supported classes, but the motivation appears largely heuristic. Why should this particular modification produce more useful OOD information than other ways of perturbing the model? Does the proposed attenuation actually weaken semantic knowledge, or does it simply change the model's predictions? The optimization forces the modified model to assign lower probabilities to locally available classes. However, this does not necessarily mean that the model has forgotten the corresponding semantic information.

7 comments
AuthorsOpenReview ↗

Response to Reviewer DGfh (Part 1)

W1: The novelty of the proposed methods is not sufficient; ... the authors consider perturbing the global model for OOD detection, which has been done in various works. The difference lies in the federated setting, and this makes the contribution of the paper incremental.

Response to W1

Our contribution is not the general idea of perturbing a trained model, but a client-conditioned intervention and response-based readout designed for Post-hoc Federated OOD Detection. In this setting, federation-level ID includes both L-ID and T-ID supported by knowledge learned across clients, whereas a deployment client retains only its L-ID data. Because local unfamiliarity does not imply OOD, the detector must separate genuine OOD from both L-ID and T-ID. OOD detection is often needed after a federated model has been trained and deployed, when retraining or further client–server communication can be costly. Unlike centralized post-hoc OOD detection, which assumes that the ID reference adequately represents the ID distribution, federated deployment must account for T-ID beyond local semantic coverage. This motivates three connected design choices.

First, Client-conditioned Support Attenuation (Sec. 4.1) uses the client label-support profile to guide the intervention. Uniformly redistributing a sample's removed true-class probability can strengthen other locally supported classes, merely shifting recognition within familiar local semantics. Our attenuation target discourages this redistribution (Eq. 1), allowing limited L-ID data to selectively weaken locally supported recognition.

Second, the representation-response readout (Sec. 4.2) examines how intermediate representations change under Client-conditioned Support Attenuation. It models multiple response patterns observed on L-ID references and evaluates whether a query follows these patterns, producing Response Alignment Support (sRs_R). Its contribution lies in interpreting attenuation-induced representation changes as ID evidence beyond changes in prediction scores.

Third, Adaptive Fusion (Eq. 5) combines Response Alignment Support (sRs_R) with Global Recognition Support (sGs_G) from the unchanged global model. Here, sRs_R provides evidence of ID-like behavior from local response references, while sGs_G provides evidence for federation-level ID beyond local semantic coverage. Each query's response controls the contribution of sGs_G, helping preserve legitimate T-ID while limiting excessive reliance on global recognition for OOD inputs.

We respectfully disagree that FedOPhD is merely an incremental extension of existing perturbation-based OOD detectors to FL. In Appendix E.2.7 (Tables 27–28), we compare classifier-weight perturbation (WeiPer+KLD), input perturbation (PRO-MSP, PRO-ENT, PRO-GEN), and intermediate-feature intervention (Gradient Short-Circuit). Rather than merely changing where perturbation is applied, FedOPhD constructs its intervention using the client label-support profile and interprets induced representation responses as ID evidence to address the mismatch between incomplete client-local ID references and federation-level ID support.

Under strong heterogeneity (α=0.1\alpha=0.1), FedOPhD improves Near-OOD AUROC/FPR95 over the strongest perturbation-based baseline (PRO-ENT) from 83.54/47.75 to 86.03/43.96 on CIFAR-10 and from 73.25/69.46 to 76.79/67.45 on CIFAR-100. For Far-OOD, compared with PRO-GEN, it improves AUROC/FPR95 from 88.14/41.67 to 97.13/13.96 on CIFAR-10 and from 79.44/60.65 to 94.06/23.80 on CIFAR-100. Across both datasets, all three heterogeneity levels, and both OOD regimes, FedOPhD achieves the highest mean AUROC and lowest mean FPR95 among all evaluated perturbation-based detectors (Tables 27–28).

Furthermore, controlled ablations (Tables 19–22) compare alternative diagnostic interventions in the full detection pipeline and Response-only settings. At α=0.1\alpha=0.1, norm-matched Random Parameter Perturbation reduces Far-OOD AUROC from 97.13 to 80.37 on CIFAR-10 and from 94.06 to 68.38 on CIFAR-100, supporting the importance of local-data-directed updates. Under the same setting, replacing Client-conditioned Support Attenuation with Uniform Redistribution changes CIFAR-100 Far-OOD AUROC/FPR95 from 94.06/23.80 to 92.59/33.62, supporting the contribution of the client label-support profile to the attenuation target.

The response-evidence and fusion analyses (Tables 12–16), together with the static-feature and update-norm controls (Tables 30–31), further evaluate the readout and fusion contributions beyond the intervention itself. These analyses are included in Appendix E.2 of the updated Rebuttal Version.

AuthorsOpenReview ↗

Response to Reviewer DGfh (Part 2)

W2: In addition, ... a hand-engineering problem, rather than a research problem.

Response to W2

We respectfully disagree that FedOPhD amounts to mere hand-engineering rather than a substantive research contribution. It addresses a specific challenge in Post-hoc Federated OOD Detection: distinguishing genuine OOD from both L-ID and T-ID using only a trained global model and current-client L-ID data. Unlike training-time federated OOD methods such as FOSTER, FOOGD, and FedSDWC, a post-hoc detector cannot acquire additional cross-client information or modify federated training.

This challenge produces two conflicting failure modes: detectors relying on incomplete L-ID references may reject legitimate T-ID, while those relying on global recognition may accept high-confidence OOD. Our L-ID-calibrated deployment evaluation (Table 6) directly illustrates this conflict. On CIFAR-10 at α=0.1\alpha=0.1, with a target L-ID false-rejection rate of 5%, FedViM rejects 52.26% of T-ID samples, while FedEnergy maintains a low T-ID false-rejection rate of 5.31% but detects only 48.93% of Far-OOD samples. In contrast, FedOPhD achieves a 7.22% T-ID false-rejection rate while detecting 92.70% of Far-OOD samples. These results demonstrate the practical importance of jointly preserving federation-level ID and detecting genuine OOD under incomplete local information.

The intervention, response readout, and fusion mechanisms address distinct aspects of this information mismatch rather than forming an arbitrary combination. Their design is supported by theoretical analysis and targeted experiments, instead of relying solely on empirical parameter choices. Theorem 4.1 establishes non-saturating excitation under attenuation, while Lemma A.1 characterizes the client-profile-directed component of the attenuation target. Proposition A.2 connects induced representation responses to client-conditioned gradients through cross-Jacobian coupling. Propositions A.3–A.4 further characterize response-gated fusion and the conditions governing T-ID preservation.

Controlled experiments further isolate the contributions of these mechanisms. Tables 19–22 compare alternative diagnostic interventions, including norm-matched Random Parameter Perturbation and Uniform Redistribution, in the full detection pipeline and Response-only settings. These comparisons support the role of local-data-directed updates and client-conditioned target construction. Table 16 and Tables 30–31 further evaluate routing, scoring, static-feature, and update-norm controls, supporting the contributions of the response readout beyond static feature geometry and the attenuation intervention beyond update magnitude.

On CIFAR-10 at α=0.1\alpha=0.1, FedOPhD-R achieves Near/Far-OOD AUROC of 64.71/91.79, while FedEnergy achieves 83.73/85.14. Combining both evidence sources through FedOPhD improves these values to 86.03/97.13 (Table 1). Furthermore, the Mean-matched Fixed Gate experiment (Table 15) isolates query dependence from average fusion weighting. On CIFAR-100 at α=0.1\alpha=0.1, the fixed gate yields Far-OOD FPR95 of 52.66%, compared with 23.80% for Adaptive Fusion. This supports the contribution of query-dependent fusion beyond selecting a favorable fixed weighting.

Moreover, without federated retraining or further client–server communication, FedOPhD's Near-OOD AUROC is within 0.85 percentage points of the strongest evaluated training-time federated method across all six main settings, outperforming that benchmark in two. For Far-OOD, FedOPhD surpasses all evaluated training-time methods on both AUROC and FPR95 across all six settings (Table 1). Our runtime evaluation (Appendix E.3.8, Table 44) further shows lower combined setup and full-inference costs than these training-time methods under the reported workloads. The deployed global classifier remains unchanged.

Taken together, the defined deployment problem, theoretical analysis, and controlled experimental evidence support FedOPhD as a substantive research contribution with practical deployment value. These analyses are included in Appendices A, E.1.1, E.2, and E.3.8 of the updated Rebuttal Version.

W3: Finally, many of the authors' sentences make the paper feel awkward and unnatural. For example: "Federated learning (FL) decouples what a client has seen from what its model has learned" ... writing this makes it difficult to understand.

Response to W3

We thank the reviewer for this feedback. While we understand that stylistic preferences may vary across readers, we have retained the original wording after careful consideration, as we believe it concisely and clearly conveys the distinction between local experience and federation-level knowledge. We would be happy to discuss further if the reviewer feels this point warrants additional clarification.

AuthorsOpenReview ↗

Response to Reviewer DGfh (Part 3)

Q1: Why is the proposed client-conditioned support attenuation necessary?

Q2: The authors design a target distribution that suppresses locally supported classes, but the motivation appears largely heuristic. Why should this particular modification produce more useful OOD information than other ways of perturbing the model?

Response to Q1 & Q2

We thank the reviewer for these questions. Client-conditioned Support Attenuation is guided by three design insights for heterogeneous FL, supported by theoretical analysis and controlled experiments.

(i) Attenuating one Local Class should avoid strengthening other Local Classes.

In heterogeneous FL, each client observes only a subset of federation-level ID classes. Uniformly redistributing the suppressed true-class probability can strengthen other locally supported classes, merely shifting recognition within familiar local semantics. Eq. (1) addresses this issue through the client label-support profile πk\pi_k:

qn=argmin⁡q∈Δ−yn[12(q−pn)⊤(q−pn)+ωnπk⊤q].q_n=\underset{q\in\Delta_{-y_n}}{\operatorname{argmin}}\left[\frac{1}{2}(q-p_n)^\top(q-p_n)+\omega_n\pi_k^\top q\right].

The first term preserves the original prediction structure, while the second discourages assigning probability to classes with stronger client-level support. Lemma A.1 characterizes the additional client-profile-aligned residual component beyond uniform redistribution.

The Uniform Redistribution control directly evaluates this design. At α=0.1\alpha=0.1, FedOPhD achieves Near-OOD AUROC/FPR95 of 86.03/43.96 versus 82.70/50.04 on CIFAR-10 and 76.79/67.45 versus 72.84/74.81 on CIFAR-100. For Far-OOD, the corresponding results are 97.13/13.96 versus 95.93/14.68 and 94.06/23.80 versus 92.59/33.62, respectively (Tables 19–20).

(ii) The probe must extract informative responses from limited L-ID data, including confidently recognized samples.

As client heterogeneity increases, local semantic coverage becomes more limited. Because the probe uses only available L-ID data, it must maintain an effective attenuation signal even when the global model already recognizes these samples confidently. Theorem 4.1 establishes

∥pn−qn∥2≥CC−1 ωn,\|p_n-q_n\|_2\geq\sqrt{\frac{C}{C-1}}\,\omega_n,

while the forward cross-entropy residual satisfies

∥pn−eyn∥2≤2(1−ωn),\|p_n-e_{y_n}\|_2\leq\sqrt{2}(1-\omega_n),

which approaches zero as ωn→1\omega_n\to1. Thus, true-class attenuation provides non-vanishing excitation on confidently recognized samples, unlike ordinary forward cross-entropy training. The role of client conditioning is distinct: the client label-support profile πk\pi_k shapes the direction of this excitation by discouraging probability redistribution toward other locally supported classes, as characterized by Lemma A.1.

The Forward Training and Norm-matched Forward controls (Tables 19–20 and 30–31) evaluate this distinction. At α=0.1\alpha=0.1, Forward Training achieves Near/Far-OOD AUROC/FPR95 of 81.17/53.53 and 91.78/22.45 on CIFAR-10, compared with 86.03/43.96 and 97.13/13.96 for FedOPhD. On CIFAR-100, the corresponding results are 71.50/77.38 and 72.49/63.87, compared with 76.79/67.45 and 94.06/23.80 for FedOPhD (Tables 19–20).

(iii) The intervention should follow local-data directions rather than random parameter directions.

Our intervention uses local adaptation samples to determine the update direction. Appendix A.2 proves that the first attenuation update lies in the local-data tangent span Uk\mathcal U_k, whereas a norm-matched isotropic perturbation places only an expected fraction rk/Pr_k/P of its squared norm in this span.

Proposition A.2 further characterizes the query response in the small-step regime:

dℓ(x)=−TηJhℓ(x;θ0)g0+O(η2),d^\ell(x)=-T\eta J_{h^\ell}(x;\theta_0)g_0+O(\eta^2),

g0=1N∑n=1NJf(xn;θ0)⊤(pn−qn).g_0=\frac1N\sum_{n=1}^{N}J_f(x_n;\theta_0)^\top(p_n-q_n).

This describes how a query's internal representations respond to attenuation guided by the client's L-ID data. FedOPhD compares these responses with L-ID reference patterns to extract ID evidence for OOD detection.

The Random Parameter Perturbation control evaluates this data-directed intervention. At α=0.1\alpha=0.1, it achieves Near/Far-OOD AUROC/FPR95 of 78.57/60.22 and 80.37/53.36 on CIFAR-10, compared with 86.03/43.96 and 97.13/13.96 for FedOPhD. On CIFAR-100, the corresponding results are 69.02/74.31 and 68.38/75.06, compared with 76.79/67.45 and 94.06/23.80 for FedOPhD (Tables 19–20).

The Static Features control further examines the value of intervention-induced representation responses. At α=0.1\alpha=0.1, it achieves Near/Far-OOD AUROC/FPR95 of 80.95/57.28 and 92.60/15.07 on CIFAR-10, and 71.20/74.88 and 87.08/43.34 on CIFAR-100, whereas FedOPhD achieves stronger results using attenuation-induced responses (Table 31). The Response-only comparisons in Tables 21–22 provide additional evidence.

These theoretical analyses and controlled comparisons are included in Appendix A and Appendix E.2 (Tables 19–22 and 30–31) of the updated Rebuttal Version.

Reviewer_5ASzOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper aims to distinguish truly unfamiliar data from normal data learned by other clients. It temporarily weakens the model’s ability to recognize local classes, observes changes in intermediate-layer features, and combines them with the original model’s predictions to detect OOD samples.

Critical strengths

I used OpenAI Codex to polish the language and improve the clarity and organization of my review. I also used it as an auxiliary tool to help identify and examine potentially inconsistent or unreasonable aspects of the paper's mathematical formulations, assumptions, and quantitative estimates. The technical assessments, critical judgments, and final recommendations are based on my own evaluation of the paper. Codex was used as an assistance tool rather than as a substitute for my independent review.

Critical weaknesses

On CIFAR-100, FedOPhD has an FPR95 of 67.45%, meaning approximately 67.45% of ID samples are incorrectly rejected when detecting at least 95% of OOD samples. This limits its reliability when false rejections are costly.

T-ID samples are assigned to other clients but may still belong to classes seen during local training. Evaluating classes entirely unseen by the receiving client would provide stronger evidence of its ability to distinguish T-ID from genuine OOD.

Questions for authors

Please see the weaknesses discussed above.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 5ASz (Part 1)

W1: On CIFAR-100, FedOPhD has an FPR95 of 67.45%, meaning approximately 67.45% of ID samples are incorrectly rejected when detecting at least 95% of OOD samples. This limits its reliability when false rejections are costly.

Response to W1

We appreciate the reviewer's concern regarding the relatively high FPR95 on CIFAR-100. This reflects the difficulty of separating overlapping ID and near-OOD score distributions at a stringent operating point requiring 95% OOD recall. As shown in Table 1, this challenge is shared by existing methods, including those requiring additional federated training. While we acknowledge the resulting false-rejection cost, it is important to distinguish this evaluation operating point from actual deployment.

Specifically, FPR95 uses an OOD-dependent evaluation threshold, whereas FedOPhD calibrates its deployment threshold exclusively on local ID data, targeting 95% L-ID acceptance. As shown in Table 6, on CIFAR-100 with α=0.1, FedOPhD achieves L-ID and T-ID false-rejection rates of 5.25% and 7.11%, respectively. Under this deployment setting, FedOPhD also achieves the highest near- and far-OOD detection rates among all evaluated post-hoc methods, outperforming FedGradPCA, the strongest post-hoc baseline in detection rate, by 2.94 and 35.27 percentage points, respectively. Notably, FedGradPCA's relatively strong detection rates come at the cost of a much higher T-ID FRR, while FedOPhD provides substantially better T-ID preservation (20.26% vs. 7.11%).

These results are particularly meaningful under FedOPhD's restrictive post-hoc setting, where only the trained global model and current-client L-ID data are available, without revisiting federated training. Despite these constraints, FedOPhD achieves the highest near-OOD AUROC among all evaluated post-hoc methods across all six main settings and the lowest near-OOD FPR95 in five. It also achieves the best far-OOD AUROC and FPR95 across all six settings, surpassing even training-time federated OOD detectors. These results demonstrate the effectiveness of FedOPhD under restricted deployment conditions. We thank the reviewer for this valuable comment and agree that further improving near-OOD detection at high OOD recall remains an important direction for future research.

We have further clarified the distinction between these evaluation and deployment operating points in Appendix E.1.1 of the updated Rebuttal Version of our manuscript.

AuthorsOpenReview ↗

Response to Reviewer 5ASz (Part 2)

W2: T-ID samples are assigned to other clients but may still belong to classes seen during local training. Evaluating classes entirely unseen by the receiving client would provide stronger evidence of its ability to distinguish T-ID from genuine OOD.

Response to W2

We thank the reviewer for this valuable suggestion. We agree that our original T-ID evaluation, defined by client provenance, may include classes already observed during the receiving client's local training. Evaluating entirely unseen classes provides a stricter test of whether FedOPhD can preserve legitimate ID learned from other clients but never observed locally.

To address this concern, we conduct an additional evaluation on CIFAR-10, CIFAR-100, and TinyImageNet under strong heterogeneity (α=0.1), using the same five random seeds. For each receiving client, we consider only T-ID samples whose classes have zero occurrences in its original FL training data, rather than merely being absent from its post-hoc reference subset. We retain the original trained models, fitted detectors, and L-ID-calibrated thresholds, with no retraining or recalibration. The L-ID and OOD evaluation populations remain unchanged. Strictly unseen T-ID false-rejection rates are aggregated using the original client weights, normalized over clients with valid unseen-class samples.

Table 11: Preservation of strictly unseen T-ID classes under α=0.1 (%, five-seed means).

DatasetValidUnseen cls.L-ID FRR↓T-ID FRR↓Unseen T-ID FRR↓Near-DR↑Far-DR↑
CIFAR-1019.2/205.455.067.227.3944.3692.70
CIFAR-10020/2050.805.257.117.2323.8673.29
TinyImageNet20/2098.505.046.707.2683.5386.53

Notes. Valid denotes the average number of receiving clients with qualifying strict-unseen T-ID samples across five seeds. Unseen cls. denotes the average number of locally unseen classes. T-ID FRR and Unseen T-ID FRR denote the original and strictly unseen T-ID false-rejection rates, respectively. DR denotes the OOD detection rate.

The results show that excluding all locally observed classes has little effect on T-ID preservation. Compared with the original T-ID evaluation, the false-rejection rate increases by only 0.17, 0.12, and 0.56 percentage points on CIFAR-10, CIFAR-100, and TinyImageNet, respectively. Across all three datasets, over 92% of strictly unseen T-ID samples are still accepted as ID. This is particularly encouraging on CIFAR-100 and TinyImageNet, where receiving clients lack an average of 50.80 and 98.50 classes, respectively, yet the strict-unseen false-rejection rates remain close to those of the original T-ID populations.

Importantly, these results are obtained at the same locally calibrated deployment thresholds, with the OOD detection rates reported in the table remaining unchanged. They provide stronger evidence for our central motivation: FedOPhD does not simply reject classes absent from local training, but can preserve legitimate federation-level ID supported by knowledge learned from other clients.

We have incorporated this stricter evaluation into Appendix E.1.4 (Table 11) of the updated Rebuttal Version of our manuscript.

AuthorsOpenReview ↗

Sincere Appreciation for Your Valuable Feedback

We sincerely thank you for your careful review, positive assessment, and valuable suggestions regarding our work.

As the author–reviewer discussion period is approaching its end, we would like to kindly ask whether our responses and additional experiments have adequately addressed your concerns. If any questions remain or further clarification is needed, we would be more than happy to provide additional explanations during the remaining discussion period.

Sincerely, we greatly appreciate your recognition and constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.