{"id":"faf0f802-8157-460a-b362-fde5a63b8ef8","arxiv_id":"2505.04979","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FedDDL improves federated out-of-distribution generalization by generating background-mixed counterfactual samples and aligning clients with causal prototypes, yielding an average Top-1 gain of about 4.5 percent over nine baselines.","lead":"Federated learning models often use background cues as shortcuts, which hurts performance on new data. This paper introduces FedDDL, which mixes backgrounds into counterfactual images and aligns client prototypes so models focus on objects, reporting about 4.5 percent higher accuracy on two NICO benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Even with a perfect object detector, the DEC module as written does not break the background-label association, so the central causal deconfounding claim is unsupported without further clarification.","rationale":"The reader's conditional verdict is appropriate, but the most load-bearing concern is not primarily the detector's reliability. Rather, the algorithm as written does not appear to implement the causal intervention it claims. Even granting Grounding DINO perfect object localization, Eqs. 5-6 generate counterfactual samples using backgrounds from the same class as the object, so the background remains predictive of the label. This is an internal inconsistency between the theoretical claim of backdoor adjustment and the concrete algorithmic recipe. The empirical gains may still be real, but they would then be better explained as a data-augmentation or style-diversification effect, not as causal deconfounding. The reader's detector concern is a valid secondary risk, and the absence of released code and the missing comparison with closely related prior work further support a conditional verdict. I would keep the verdict as conditional, with the additional requirement that the authors either specify cross-class background sampling or reframe the contribution on non-causal grounds and provide evidence that the background-label association is actually reduced.","tokens_in":14212,"tokens_out":5558,"duration_ms":62157,"concrete_test":"Estimate the background-label association in the generated counterfactual set D_C,k under Eqs. 5-6. Concretely, for each class i, record the class identity of the source backgrounds used to create each counterfactual sample and compute the mutual information I(B;Y) or the accuracy of a background-only classifier trained to predict the object's label from the background pool used in D_C,k. If I(B;Y) remains comparable to the original training set, the DEC module does not implement the P(B) intervention of Eq. 2. Then re-run the NICO-Animal (L7) experiment with a modified DEC that explicitly pools backgrounds across all ten classes before grouping, while keeping every other setting identical. If accuracy does not change materially, the reported improvement is not attributable to deconfounding.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that FedDDL performs backdoor adjustment to make background B independent of label Y. This requires the counterfactual training distribution to have B and Y independent. But the algorithm as specified in Eqs. 5-6 does not achieve that. Eq. 5 splits only the backgrounds I_i^B 'belonging to class i' into groups, and Eq. 6 fuses those same-class backgrounds with an object I_O. If I_i^B is class i's own background set, then every generated counterfactual sample preserves the original background-label correlation: dogs still appear with dog-class backgrounds, cows with cow-class backgrounds. The text claims the counterfactual samples 'establish a connection between the background and any label,' which would require sampling backgrounds from all classes and pairing them with objects of every class, but that operation is not described. Eq. 7 further assigns the original true class label to each counterfactual sample, so the background-label co-occurrence in D_C,k mirrors the original training data. Consequently, the reported accuracy gains may come from intra-class background diversification and data augmentation rather than from the claimed causal deconfounding. The reader's identified weakness about Grounding DINO reliability is legitimate, but it is secondary: even a perfect detector cannot produce the claimed intervention if the counterfactual construction does not randomize backgrounds across labels. The paper should either specify cross-class background pooling explicitly or revise the causal claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes FedDDL, a federated learning method for out-of-distribution generalization. The authors build a structural causal model in which image background B and data source S act as confounders of the image X and label Y, and they claim that an intra-client deconfounding module (DEC) implements backdoor adjustment by generating counterfactual images with recombined backgrounds, while an inter-client debiasing module (DEB) aligns client representations through class-wise causal prototypes computed from object-only images under a contrastive loss. The method is evaluated on NICO-Animal and NICO-Vehicle in terms of global and local Top-1 accuracy against nine baselines, with ablations, hyperparameter sensitivity studies, and qualitative attention analyses; the reported average improvement over the best baseline is about 4.5 percentage points.","tokens_in":14507,"tokens_out":10470,"duration_ms":100373,"significance":"The empirical component of the paper is solid: the evaluation uses standard benchmarks with external test sets and held-out backgrounds, nine baselines, three-run statistics, a module-level ablation (Table 3), hyperparameter sensitivity (Fig. 6), a plug-and-play demonstration (MOON+DEC in Table 2), and honest reporting of local-model variance. The global-model gains are large relative to the reported standard deviations, and the claimed average improvement of roughly 4.5% over the strongest baseline is consistent with Table 2. If the deconfounding interpretation were correct, this would be a principled new perspective for federated OOD generalization. However, as specified, the DEC module is an intra-class background-mixing data augmentation and the DEB module is a prototype-based contrastive regularizer; the paper does not establish that either procedure implements the backdoor adjustment in Eqs. (2) and (9). The causal narrative is therefore not supported by the described algorithm, and the durable contribution is a well-evaluated empirical method whose causal interpretation needs either a corrected specification or a softened claim.","major_comments":[{"comment":"The counterfactual generation procedure as written does not randomize backgrounds across labels, so it does not produce the interventional distribution P(Y|do(X)) claimed in Eq. (2). Eq. (5) groups backgrounds I_i^B 'belonging to class i' into class-specific groups g_j^i, and Eq. (6) fuses those same-class background groups with the object I_O; the classification loss in Eq. (7) then assigns the original true class label to each counterfactual sample. Under the most direct reading, every counterfactual is a class-i image with a class-i background, so the background-label co-occurrence in D_C,k mirrors the training data and the spurious B→Y path is preserved. The abstract's claim that the counterfactuals 'establish a connection between the background and any label' would require pairing background groups sampled from all classes with objects of all classes, but that operation is not described in the method text or in Algorithm 1. Please specify the cross-class pairing rule (including the label rule for mixed pairs), or alternatively revise the causal-intervention claim.","section":"§4.2, Eqs. (5)–(7)"},{"comment":"Even under the intended cross-class pairing, the paper gives no argument that the generated counterfactual distribution matches Σ_B P(Y|X,B)P(B). The Split(·) operator in Eq. (5) creates η class-indexed groups, and Eq. (6) averages the images within each group, so the counterfactual backgrounds are class-conditional averages rather than samples from the background marginal P(B). In addition, the two terms of Eq. (7) are added with equal weight, which does not correspond to any stated weighting by P(B). Without an explicit account of how the counterfactual sampling distribution relates to the intervention in Eq. (2), the equality connecting the backdoor formula to the DEC loss is not established.","section":"§4.2, Eqs. (2) and (6)"},{"comment":"The DEB module is described as performing the interventional computation in Eq. (9), but the module actually computes class-wise prototypes from object crops (Eq. 10), averages them across clients (Eq. 11), and applies a supervised contrastive loss (Eq. 12). No term in L_CR corresponds to the summation over P(B)P(S_i) in Eq. (9); P(S_i) is never defined for the fixed, finite set of FL clients, and the text does not explain why aligning local features to object-based prototypes instantiates the backdoor adjustment. The module's actual effect—cross-client representation alignment—is a legitimate technical goal and should be claimed directly rather than through Eq. (9).","section":"§4.3, Eqs. (9)–(12)"},{"comment":"The entire method rests on Grounding DINO's bounding boxes, with the BOX THRESHOLD and TEXT THRESHOLD fixed at 0.3, but the paper reports no measure of detection quality (e.g., per-class detection rate or object/background separation quality) and no sensitivity analysis over these thresholds. If the detector misses small or occluded objects, then I_B still contains the causal object and I_O can be near-empty, which would corrupt both the counterfactual samples of Eq. (6) and the causal prototypes of Eq. (10). A quantitative check of the detector's reliability is needed to support the claim that the object/background decomposition is sound.","section":"§5.1, Eqs. (3)–(4)"}],"minor_comments":[{"comment":"The second term of Eq. (7) is garbled ('+ −Σ_{C}^{n=1} y_C(n) log(...)'); it should be written over a counterfactual-sample index with a consistent summation bound, and the relative weight of the two loss terms should be stated explicitly.","section":"Eq. (7)"},{"comment":"Line 18 aggregates θ^t as the average of the previous-round parameters θ^{t−1}_k rather than the updated local models θ^t_k, and the pseudocode's uniform 1/|K| weighting is inconsistent with the dataset-size weights α_k defined in Section 3; please clarify which aggregation rule was used in the experiments.","section":"Algorithm 1, lines 16–19"},{"comment":"The Figure 3 caption states that the modules sever 'B ↛ X' and 'S ↛ X', while Section 4.1 says the modules sever 'X ↚ B' and 'X ↚ S' and the path B→Y; the target of the claimed intervention should be stated consistently throughout.","section":"§4.1 and Figure 3 caption"},{"comment":"The baseline name appears as 'FedDeccor' in Table 2 but 'FedDecorr' in Section 5.2; please make the spelling consistent.","section":"Table 2"},{"comment":"The phrase 'decbiasing learning module' appears twice in the ablation discussion and should read 'debiasing learning module'.","section":"§5.3"},{"comment":"The text contains an unresolved placeholder citation, 'Following prior studies [?; Liu et al., 2021]'; the missing reference should be supplied.","section":"§5.1, Implementation Details"},{"comment":"Several entries use the nonstandard form '[Qi and et al., 2024]' and '[Zhang and et al., 2024]' with the author list replaced by 'and et al.', and the names in a few entries (e.g., '´Slazyk') contain stray accent artifacts; the bibliography should be brought into the venue's format.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The causal framing is the main editorial risk: the operations actually described in the paper are within-class counterfactual augmentation and prototype-based contrastive learning, and the backdoor-adjustment interpretation is not realized by Eqs. (5)–(7) and (9)–(12). If the authors can either correct the counterfactual specification to genuinely mix backgrounds across labels or scale the claims down to data augmentation plus representation alignment, the paper becomes publishable on the strength of its empirical study. The citation pattern is heavily weighted toward the authors' own prior work (multiple Qi et al. entries in Sections 1–2), which is understandable given their prior line of research but should be checked for even-handedness. The empirical evaluation is stronger than typical for this area, so a revise-and-resubmit path is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: FedDDL is a plausible empirical recipe, but the paper's central causal claim does not survive contact with its own equations. The method—background/object decoupling via Grounding DINO, counterfactual samples, and prototype contrastive alignment—looks like a useful bag of tricks for federated OOD generalization, and the reported gains are consistent across ablations. But if you read Eqs. 5–6 carefully, the counterfactual construction only mixes backgrounds within a class, so the background-label correlation is never broken. The claim that the model stops using background to infer the label is not supported by the described procedure.\n\nWhat is actually new: the specific combination of intra-client deconfounding (DEC) and inter-client debiasing (DEB) in one FL pipeline, with object-only prototypes for the contrastive term. The empirical protocol is standard: two NICO splits, nine baselines, ablations, hyperparameter sensitivity, and attention visualizations. That is real work, and the numbers look plausible. Giving credit where it is due, the ablations monotonically improve and the plug-and-play result (MOON+DEC) is a nice sanity check.\n\nThe soft spots are proportionate to the paper's ambition. First and most importantly, the causal derivation is not just loose; it is wrong as written. Eq. (2) and Eq. (9) are backdoor adjustment formulas, but the algorithm never performs the required intervention. The stress-test note is correct: even a perfect detector cannot fix this, because the same-class background pooling preserves spurious association. Second, the closest prior work from the same group (Zhang et al., 2025b) is cited but not compared, which is a visible gap. Third, no code is released, so the empirical results are hard to verify.\n\nOne minor note: the DINO detection quality issue flagged by the reader is legitimate but secondary. The bigger problem is upstream.\n\nWho should read it: anyone working on federated OOD generalization can learn from the empirical setup and the negative lesson about aligning causal language with actual augmentation. It deserves referee time—an editor should send it out—but the authors need to either fix the counterfactual construction to truly pool backgrounds across classes or drop the causal framing and claim domain randomization.\n\nMy peer-review recommendation: engage with it, but with the expectation of major revision.","headline":"A plausible federated OOD recipe whose central causal deconfounding claim does not survive contact with its own equations; deserves peer review with major revision, not desk rejection.","tokens_in":15012,"tokens_out":3346,"would_cite":false,"duration_ms":32587,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FedDDL claims that removing background and source confounders from federated inference raises out-of-distribution accuracy by 4.5% on average.","keywords":["federated learning","out-of-distribution generalization","deconfounding","backdoor adjustment","counterfactual samples","causal prototypes","background bias","object detection"],"falsifier":"Train FedDDL on a version of NICO where the object detector is deliberately weakened (e.g., small, occluded, or camouflaged objects with lower detection confidence): if the 4.5% average gain collapses below FedAvg's accuracy while all other components remain unchanged, the claimed deconfounding is not what drives the improvement.","tokens_in":1840,"feed_emoji":"🎯","tokens_out":2190,"duration_ms":75802,"temperature":0.7,"pith_summary":"The paper argues that federated models fail on out-of-distribution data because they learn two spurious inference paths: a background factor $B$ that correlates with labels within each client, and a source factor $S$ that differs between clients. It proposes FedDDL, which builds a structural causal graph over these factors and performs backdoor adjustment to block both paths. An intra-client module splits each image into object and background with a text-prompted detector, then fuses the object with random backgrounds so the model cannot lean on scenery. An inter-client module constructs prototypes from object images alone and aligns local features to them, shrinking cross-client output gaps. If the causal repair works as described, it explains the reported 4.5% average Top-1 accuracy improvement over nine baselines on NICO-Animal and NICO-Vehicle.","feed_headline":"Causal repair lifts federated accuracy by 4.5%","feed_subtitle":"Cutting background and source confounders out of federated models lets them focus on objects in unseen scenes.","key_machinery":"The load-bearing machinery is the structural causal graph over Source $S$, Background $B$, Object $O$, Image $X$, and Label $Y$, together with the backdoor adjustment formula $P(Y|\\mathrm{do}(X)) = \\sum_B P(Y|X,B)P(B)$. The mechanism runs through a pre-trained open-set object detector (Grounding DINO) prompted with the class name to produce a bounding box; Hadamard masks split each image into object image $I_O$ and background image $I_B$; counterfactual samples $I_C$ are created by fusing $I_O$ with randomly grouped backgrounds from the same client; and causal prototypes are class-wise mean embeddings of $I_O$ computed by the global model, used in a contrastive loss that pulls local features toward same-class prototypes and away from other-class prototypes. All of this is designed to make the background contribute equally across labels and to align heterogeneous client representations so that aggregation does not carry spurious correlations.","core_discovery":"The central claim is that federated out-of-distribution generalization fails through two identifiable confounding paths, $X \\leftarrow B \\to Y$ and $X \\leftarrow S \\to Y$, and that both can be severed without exchanging raw samples. FedDDL replaces the confounded inference $P(Y|X)$ with the backdoor-adjusted $P(Y|\\mathrm{do}(X))$ by generating counterfactual images whose backgrounds are decoupled from labels and by regularizing local representations against causal prototypes computed from object regions alone. The result is that models learn to predict from the object feature $O$, so unseen backgrounds no longer hijack the decision. The paper presents this as both a causal analysis of the inference paths and a method, with the 4.5% average Top-1 gain as evidence that the deconfounding, not mere data augmentation, drives the improvement.","pith_inferences":["If the causal story is right, FedDDL should also improve robustness to background shifts not present in NICO's seven training backgrounds, for example by evaluating on held-out scenes from other datasets, which the paper does not test.","Because counterfactual backgrounds are mixed only within each client's own images, the method preserves the privacy boundary of federated learning more strictly than style-transfer augmentation that exchanges style statistics across clients; this is an implied benefit, not a claim the paper tests.","A natural extension, which the paper names as future work, is replacing the fixed text-prompted detector with learned or adaptive object localization when attributes cannot be directly disentangled, such as in medical images.","The method's object-background split is binary; extending it to multi-part object structure (e.g., animal head versus body versus habitat) would test whether the deconfounding gain comes from removing background only or from isolating the most discriminative sub-region."],"forward_implications":["FedDDL reports 4.5% higher Top-1 accuracy on average than nine baselines across NICO-Animal and NICO-Vehicle, improving both global and averaged local models.","The DEC module works as a plug-in: adding it to MOON (MOON+DEC) improves that baseline substantially, suggesting the counterfactual intervention transfers across federated learning methods.","The DEB module reduces performance disparity among local clients by aligning their representations to shared causal prototypes, as shown by lower local accuracy variance and t-SNE alignment.","Ablation on the L7 splits shows the gains of DEC and DEB are additive: full FedDDL exceeds either module alone, and benefit grows with the number of counterfactual background groups $\\eta$.","With a fixed ResNet-18 backbone, FedDDL remains above FedAvg across the tested hyperparameter ranges $\\lambda \\in \\{0.1, 1.0, 2.0\\}$, $\\tau \\in \\{0.07, 0.5\\}$, and $\\eta \\in \\{1, 3, 5\\}$."],"supporting_citations":[{"why":"It supplies the open-set detector, Grounding DINO, used in Eq. 3-4 to split each image into object and background regions.","marker":"[Liu and et al., 2025]"},{"why":"It provides the NICO-Animal and NICO-Vehicle datasets whose background-based train/test splits define the out-of-distribution evaluation.","marker":"[Wang et al., 2021]"},{"why":"It provides the FedAvg baseline and aggregation protocol that FedDDL builds on and compares against.","marker":"[McMahan et al., 2017]"},{"why":"It provides the MOON baseline and the target for the DEC plug-in experiment (MOON+DEC).","marker":"[Li et al., 2021]"},{"why":"It supplies FedCCRL, the strongest prior baseline whose reported accuracy FedDDL exceeds on all splits.","marker":"[Wang and Tang, 2024]"},{"why":"It supplies FPL, the prototype-based federated learning baseline that FedDDL's causal prototypes extend.","marker":"[Huang et al., 2023]"},{"why":"It provides the ResNet-18 backbone used uniformly across all compared methods, making the accuracy comparisons fair.","marker":"[He et al., 2016]"}],"fun_headline_variants":["Causal deconfounding lifts federated OOD by 4.5%","Severing confounders boosts federated accuracy 4.5%","FedDDL: Causal backdoor fix for federated OOD","Backdoor adjustment boosts federated OOD accuracy 4.5%","Causal prototypes cut confounders: FedDDL +4.5% OOD"],"cache_read_input_tokens":17152,"weakest_assumption_plain":"The method stands on the assumption that the text-prompted object detector reliably separates the truly causal object from the background in every client's images, so the counterfactual backgrounds and causal prototypes carry no hidden spurious information.","fun_headline_variants_meta":{"raw":{"variants":["Causal deconfounding lifts federated OOD by 4.5%","Severing confounders boosts federated accuracy 4.5%","FedDDL: Causal backdoor fix for federated OOD","Backdoor adjustment boosts federated OOD accuracy 4.5%","Causal prototypes cut confounders: FedDDL +4.5% OOD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000735,"raw_usage":{"total_tokens":3305,"prompt_tokens":981,"completion_tokens":2324,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":2222}},"tokens_in":597,"tokens_out":2324,"duration_ms":17039,"temperature":1.0,"reasoning_tokens":2222,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:15:46.345287+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train FedDDL on a version of NICO where the object detector is deliberately weakened (e.g., small, occluded, or camouflaged objects with lower detection confidence): if the 4.5% average gain collapses below FedAvg's accuracy while all other components remain unchanged, the claimed deconfounding is not what drives the improvement.","supporting_citations":[{"cited_title":"Grounding dino: Marrying dino with grounded pre-training for open-set ob- ject detection","cited_arxiv_id":null,"evidence_quote":"It supplies the open-set detector, Grounding DINO, used in Eq. 3-4 to split each image into object and background regions."},{"cited_title":"Causal attention for unbiased visual recognition","cited_arxiv_id":null,"evidence_quote":"It provides the NICO-Animal and NICO-Vehicle datasets whose background-based train/test splits define the out-of-distribution evaluation."},{"cited_title":"Communication-efficient learn- ing of deep networks from decentralized data","cited_arxiv_id":null,"evidence_quote":"It provides the FedAvg baseline and aggregation protocol that FedDDL builds on and compares against."},{"cited_title":"Re- thinking federated learning with domain shift: A prototype view","cited_arxiv_id":null,"evidence_quote":"It supplies FPL, the prototype-based federated learning baseline that FedDDL's causal prototypes extend."}],"review_version":1}