acceptodds
Under review as a conference paper at ICLR 2027

From Clients to Samples: Tracing Data Influence in Federated Learning

Abstract

Understanding how individual training data shape model behavior is essential for error diagnosis and data curation, particularly in medical applications. In Federated Learning (FL), tracing training-data influence is particularly challenging because data remain decentralized, and requires separating the effects of decentralized client updates from those of individual local samples. Existing methods primarily focus on quantifying client-level contributions, offering limited insight into the individual samples that shape their local models. Such aggregate assessments can mask samples with heterogeneous or opposing effects on model behavior, making targeted data inspection difficult. To address these challenges, we propose Federated Hierarchical Influence Tracing (FedHIT), an attribution framework for medical FL that traces influence at two complementary levels. Level 1 attributes query-specific global model behavior to individual clients' local updates, identifying the clients most responsible for performance on a given query. Level 2, applied within a client identified at Level 1, attributes that client's local model behavior to its own training samples, identifying which examples drive its local model. This decomposition connects global behavior to responsible clients and, separately, a client's local behavior to responsible samples, without requiring influence to be traced end-to-end from individual samples to the global model. We validate FedHIT on three benchmark datasets, demonstrating that it reliably identifies influential clients and samples at each level, enabling targeted inspection of the data underlying client- and global-model behavior in federated medical settings. By restricting sample-level analysis to selected clients, FedHIT avoids exhaustive attribution over all decentralized samples, providing a computationally efficient approach to targeted data inspection and curation.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Federated learning, Data influence estimation, Client contribution, Sample-level influence, Medical 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 4 reviews, above 27% of the others.
Reviewer_gZhTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FedHIT, a two-stage workflow for query-specific diagnosis in federated learning. It first ranks client updates in each communication round by their alignment with the query-loss gradient, then applies SOURCE to the local training trajectories of selected clients to identify influential samples. The two stages explain different models: client scores concern the round-local global update, while sample scores concern the selected client's local endpoint. The paper evaluates client-ranking agreement, sample attribution, removal counterfactuals, and attribution cost on three medical image datasets.

Critical strengths
  1. The distinction between the two attribution targets is explicit and useful. Sections 2.1 and 3 carefully avoid interpreting a local sample score as that sample's influence on the final global model. The removal surrogate and the sign convention are also explained clearly, which helps make the output interpretable.
  2. The client-level results are encouraging within the studied setting. Section 4.1 reports mean rank correlations of 0.936–0.988 on Matek-19, with similarly strong results on the other datasets in Appendix G.1. Table 1 provides a complementary directional removal check. The cost analysis also includes trajectory replay and distinguishes measured runtime from FLOPs rather than reporting theoretical savings alone.
Critical weaknesses
  1. The evidence for the hierarchy itself is less convincing than the evidence for its individual components. Level 2 directly applies SOURCE (Section 3), and Section 4.2 deliberately evaluates it independently of client routing. In the cost comparison, both methods expand three clients per round; the main difference is SOURCE versus explicit sample-wise retraining (Section 4.4 and Appendix D.1). This establishes an efficient implementation relative to a costly reference, but does not establish how much client routing improves the diagnostic result. For example, there is no comparison with randomly routed clients or with SOURCE applied to all clients at a matched inspection budget, nor a quantitative data-curation or error-diagnosis outcome. Given that combining the two levels is the main contribution, this missing evaluation substantially limits the case for the method.
  2. The sample-level evidence and reporting leave considerable uncertainty. Figure 3 gives mean LDS values of only 0.078–0.162, without error bars or variation over training/partition seeds. Appendix G.2's removal table is explicitly a selected illustrative query-round, so it does not substantiate the stated consistency across queries and rounds. In addition, the paper does not specify actual learning rates, optimizer and batch size, segment/checkpoint settings, curvature sample cap, or key attribution-baseline settings. These omissions make it difficult to assess whether the improvements are reliable and whether the comparison can be reproduced.
Questions for authors
  1. Could you quantify what routing contributes using the same SOURCE implementation with top-score, random, and broader/all-client routing, preferably reporting a concrete diagnosis or curation outcome at comparable cost? A routing-budget sweep would also clarify the quality/cost trade-off.
  2. Can you report query/round-aggregated sample-removal results and uncertainty across independent training or partition seeds, and provide the training, segmentation, curvature, and baseline hyperparameters? For LDS, which client-round trajectories are evaluated, what query measurement is used, and how are those results aggregated?
  3. How is the Level-1 score related to the renormalized LOCO intervention when client weights differ? Writing u = sum_i alpha_i Delta_i, first-order expansion at the round start gives LOCO loss change approximately [s_i + alpha_i g_q^T u]/(1-alpha_i), whereas the score is s_i = -alpha_i g_q^T Delta_i. Equal weights preserve the ordering under this approximation, but unequal weights need not. Please clarify whether the reported partitions have equal client sizes and whether ranking agreement persists for unequal weights.
  4. Please reconcile the local sample-score definition in Equations (2)–(3) and Algorithm 1, which uses 1/n_c, with Equation (5), which introduces alpha_c/n_c. That extra factor does not affect within-client rankings, but it changes the scale of the claimed local removal estimate. Which version is implemented?
Reviewer_QNwXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to trace the influence of training data on the behavior of a federated learning model, with a focus on medical image classification. The authors propose FedHIT, a two-stage framework. The first stage identifies which clients’ updates contributed most to the model’s behavior on a given test example. The second stage identifies which training samples within the selected clients shaped their local models. The method is evaluated on three medical imaging datasets. The authors report that it agrees with removal-based reference checks, outperforms existing sample attribution methods, and costs much less than exhaustive retraining.

Critical strengths
  1. The paper is interesting and nice to read. Tracing client and data contirbutions is interesting to study and the paper has some interesting results.
  2. The experiemental evaluation is extensive and well supported.
Critical weaknesses
  1. My main concern centers around the premise from a privacy persepctive. FL promises privacy which should mean that individual data points and client contributions should not be traceable, especially in the medical domain. Yet this paper wants to find individual data influences, which seems to be at odds with the privacy requirements of FL. So I am not sure how practical this approach is, espeically since the application to domain is medical images.
  2. My second concern is more technical. Level 1 is a first-order estimate of removing one client’s update from the aggregate, and the reference it is validated against (round-local LOCO) computes that same removal exactly. So a high Spearman correlation that increases with the number of clients, where each client’s weight shrinks and the linearization gets better, mostly confirms the approximation and says little about real influence on training. Table 1 only compares against random removal, so it is unclear whether simple scores such as update norm, cosine similarity to the query gradient, or how much of the query’s class a client holds would do just as well, which seems plausible given the label-skewed partitions. Level 2 is SOURCE applied as is to a local trajectory, and its scores are explicitly not propagated to the global model. The two levels are therefore connected only by running one after the other, and nothing in the paper tests whether a sample flagged at Level 2 matters for the global behavior that motivated the query.
Questions for authors
  1. Can the authors clarify the threat model (who runs FedHIT, who is trusted) and say clearly whether the method applies under secure aggregation or user-level DP? Do the per-sample scores that clients return to the server leak information about their training data?
  2. Can the authors add simple Level 1 baselines such as update norm, cosine similarity to the query gradient, or class share, instead of only random removal? How does the Spearman correlation vary per round?
  3. Is there evidence that Level 2 scores say something about the global model? For example, remove the top-attributed samples at a routed client-round, re-aggregate, and measure the change in the global query margin.
  4. Can the authors compare against SOURCE applied to all clients, to separate the effect of routing from the effect of the estimator?
Reviewer_rDdTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedHIT, a query-conditioned framework for hierarchical attribution in federated learning. Level 1 ranks client updates at each communication round using their alignment with the query-loss gradient. Level 2 applies SOURCE to estimate training-sample influence within selected clients’ local training trajectories. The two levels address distinct attribution targets, and the local sample scores are not propagated to the final global model. Experiments on three medical imaging datasets evaluate client-ranking agreement, sample attribution, removal effects, and computational overhead.

Critical strengths
  1. Level 1 is well supported within its stated scope. Its mean Spearman correlations with round-local LOCO range from 0.936 to 0.996 across three datasets and multiple federation sizes. The client-removal experiments provide additional evidence that the scores distinguish updates that help the query from those that hurt it.
  2. The paper clearly distinguishes the two attribution targets. In particular, it explicitly states that Level 2 explains the selected local model rather than estimating sample influence on the final global model. This distinction is important for interpreting the results and keeps the claims appropriately scoped.
Critical weaknesses
  1. The methodological contribution appears limited, and the value of the proposed combination is not sufficiently demonstrated. Level 1 uses a straightforward first-order alignment between weighted client updates and the query-loss gradient, while Level 2 directly adopts SOURCE. The main contribution therefore lies in organizing these components into a hierarchical diagnostic workflow. However, the experiments do not clearly establish what this organization adds. Section 4.2 evaluates SOURCE independently of routing, which supports the choice of the sample estimator but does not demonstrate the benefit of the combined workflow. In the efficiency comparison, both pipelines expand three clients. Appendix D.1 explicitly states that their Level-2 cost difference comes from the attribution estimator. Thus, the substantial speedups mainly show the benefit of replacing per-sample retraining with SOURCE. Although the strong LOCO agreement supports client localization, the additional diagnostic benefit of routing remains unclear without a comparison using the same estimator and attribution budget.
  2. The evidence for reliable sample-level diagnosis is not sufficiently specified or aggregated. Section 4.2 does not clearly identify which federation sizes and local client trajectories are evaluated, how subset retraining is initialized, how long it runs, or how randomness and the query measurement are handled. Important settings, including the optimizer, learning rate, batch size, SOURCE segmentation and checkpoints, and baseline configurations, are also missing. The manuscript provides no implementation or code repository link that would allow these details to be checked. Moreover, Figure 3 reports mean LDS values without uncertainty, while Table 5 presents one selected query-round example despite claims of consistent removal behavior across queries and rounds. SOURCE obtains the highest reported LDS, but the current results do not establish how stable its advantage and local diagnostic effects are.
Questions for authors
  1. What additional diagnostic benefit does the hierarchical workflow provide beyond applying SOURCE to a limited set of clients? Can you compare query-conditioned routing with random client selection while keeping SOURCE and the attribution budget fixed?
  2. Which federation sizes and client-round trajectories are used for LDS? Please clarify the subset-retraining initialization, training horizon, randomness, query measurement, and score-sign convention, and provide the training and attribution configurations needed to reproduce the comparisons.
  3. Can you report aggregate sample-removal effects and their variability across the evaluated queries and client-rounds, together with uncertainty for the LDS comparisons? These results would help assess the claimed consistency beyond the selected example in Table 5.
Reviewer_unu3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FedHIT, a two-level post-hoc attribution workflow for Federated Learning (FL) evaluated on medical imaging benchmarks. Given a test query z_q at communication round r, Level 1 scores each client's update via a first-order inner product with the server query gradient, sᵢ^((r))(z_q) = -αᵢ^((r))(Δθᵢ^((r)))^⊤∇θℓ(z_q; θ^((r))), and routes the top-k clients to Level 2. Level 2 replays the selected client's local training trajectory and directly applies SOURCE (Bae et al., 2024) to estimate sample-level influence on that client's local model endpoint ϑ(c, T)^((r)). Experiments with ResNet-18 on three medical datasets (HAM10000, Matek-19, BreakHis) verify Level 1 against round-local Leave-One-Client-Out (LOCO) and Level 2 against IF, TracIn, and TRAK.

Critical strengths

Filtering K clients down to k via server-side update-gradient inner products before running heavy trajectory-based sample attribution (SOURCE) is a sensible engineering design that keeps raw data on-device and avoids running EK-FAC across all clients.

Critical weaknesses
  1. Minimal Technical Novelty: The method is a direct sequential stitching of two existing techniques: a standard first-order Taylor inner product at Level 1, followed by an unmodified call to SOURCE (Bae et al., 2024; Eqs. 2–3 are taken directly from SOURCE) on a 3-epoch local trajectory at Level 2.
  2. Inconsistent Attribution Targets Between Level 1 and Level 2: Level 1 routes client cc based on how its update Δθc(r)\Delta\theta_c^{(r)} aligns with the server pre-aggregation gradient gq(r)=∇θℓ(zq;θ(r))g_q^{(r)} = \nabla_\theta \ell(z_q; \theta^{(r)}), whereas Level 2 computes sample influence projected onto the unaggregated local endpoint gradient ∇ϑf(zq;ϑc,T(r))\nabla_\vartheta f(z_q; \vartheta_{c, T}^{(r)}) (Eq. 3 / Alg. 1 Line 25)—and neither level propagates influence across subsequent FL rounds (r→Rr \to R). Under non-IID client drift (80:20 / 90:10 skew), ∇ϑf(zq;ϑc,T(r))≠gq(r)\nabla_\vartheta f(z_q; \vartheta_{c, T}^{(r)}) \neq g_q^{(r)}, so the two levels do not decompose a common objective even within a single round.
  3. Near-Tautological Level 1 Validation & Strawman Efficiency Baseline:
    • Level 1 vs. LOCO: Removing client ii from FedAvg perturbs θ(r+1)\theta^{(r+1)} by αi(r)1−αi(r)(Δθi(r)−Δθ(r))\frac{\alpha_i^{(r)}}{1-\alpha_i^{(r)}}(\Delta\theta_i^{(r)} - \Delta\theta^{(r)}). By Taylor expansion, round-local LOCO is analytically proportional to si(r)(zq)s_i^{(r)}(z_q) as αi(r)=1/K→0\alpha_i^{(r)} = 1/K \to 0, which trivially explains why the Spearman correlation approaches 1.01.0 as KK increases from 1010 to 5050 (0.936→0.9880.936 \to 0.988).
    • Inflated Speedups (Tables 2–3): The claimed 70%–99%70\%\text{–}99\% GFLOP reduction compares FedHIT (using SOURCE at Level 2) against brute-force per-sample LOO retraining (Θ(nc2)\Theta(n_c^2)). As Table 3 shows, >99.9%>99.9\% of the baseline cost comes from brute-force LOO retraining; thus, the reported speedup mostly reflects SOURCE vs. LOO retraining (Bae et al., 2024), whereas Level 1 routing itself only saves a factor of K/kK/k.
Questions for authors
  1. Isolating the Routing Speedup: What is the exact GFLOP and wall-clock speedup of FedHIT (k=3k=3 + SOURCE) against running SOURCE on all KK clients (rather than against brute-force per-sample LOO retraining)?
  2. Gradient Mismatch in Eq. 3: If Level 1 selects clients via −αc(Δθc(r))⊤gq(r)-\alpha_c (\Delta\theta_c^{(r)})^\top g_q^{(r)}, why doesn't Level 2 project the local segment perturbations onto the same server gradient gq(r)g_q^{(r)} (or ∇θℓ(zq;θ(r+1))\nabla_\theta \ell(z_q; \theta^{(r+1)})) instead of the drifted local endpoint gradient ∇ϑf(zq;ϑc,T(r))\nabla_\vartheta f(z_q; \vartheta_{c, T}^{(r)})?
  3. Full-Pipeline Counterfactual: If you remove the top positively/negatively attributed samples identified by FedHIT and retrain the full 200-round FL model from scratch, does the final global model θ(R)\theta^{(R)} exhibit a larger prediction shift on zqz_q than removing samples ranked by TracIn or TRAK?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.