Skip to content

Interpretable Federated Learning via Global-Local Attribution Alignment

Explainable Federated Learning via Global-Local Attribution Alignment

arXiv: 其他 AI 安全论文· arXiv:2611.20445· ICML 2026· Dawood Wasif, Terrence J. Moore, Chang-Tien Lu, Jin-Hee ChoDepartment of Computer Science, Virginia Tech; DEVCOM Army Research LaboratoryIndexed by this site · 2 days ago· Originally publishedAlignment· Its place in the scan record →

Quick read

Interpretability

Based on the paper PDF (AI-generated); refer to the original

The authors propose a federated learning explanation framework, xFedAlign, which they say maintains FedAvg accuracy on 6 benchmarks while reducing cross-client explanation drift (EDI).

Questions
In federated learning, client data is heterogeneous, local explanations drift from one another, global aggregation loses fine-grained structure, and sharing gradients or dense attribution maps carries privacy and bandwidth costs. The authors aim to obtain locally faithful and globally consistent explanations under strict privacy and bandwidth constraints.
Methods
Proposes xFedAlign: each client trains a lightweight surrogate model via distillation, produces per-class top-k attribution artifacts in a normalized group space, and after clipping, quantization, and noise addition, the server aggregates them via coordinate-wise median into a Global Explanation Prior, then a JSD alignment loss softly constrains each client's explanations, without affecting task optimization.
Experiments and results
On the IID and non-IID partitions of six benchmarks—MNIST, CIFAR-10, AG News, IMDb, Adult, and Credit—the authors state that xFedAlign matches FedAvg in accuracy and has the lowest EDI in most cases; on MNIST, membership inference accuracy and ∆Top@128 and ∆EDI under attribution poisoning are all lowest; when scaled to 128 clients, EDI stays at 0.0002.
Limitations and future work
The authors state that the framework assumes the group space is meaningful and relatively stable, and that adaptive grouping is needed when groups are ambiguous or semantics drift; the privacy analysis is limited to an honest-but-curious server and client-level DP under secure aggregation, and sparse artifacts may still leak structured information. Future directions include regression tasks, federated language models, adaptive attackers, and expert evaluation.
Experimental setupSmall CNN (vision: MNIST, CIFAR-10)、TextCNN (text: AG News, IMDb Reviews) et al. · MNIST、CIFAR-10 et al. · Accuracy、EDI et al.
Threat model
The task model is treated as a black box: clients can query predictions but do not share raw data, gradients, logits, or internal activations; the server is honest-but-curious and only sees securely aggregated, noise-added artifacts; malicious clients can submit manipulated explanation artifacts to perform attribution poisoning; the membership inference attacker is black box and can see predictions and the broadcast prior, and the strongest variant can also see local explanations.
Evaluated models
Small CNN (vision: MNIST, CIFAR-10)TextCNN (text: AG News, IMDb Reviews)Shallow MLP (tabular: UCI Adult Income, UCI German Credit)
Benchmarks
MNISTCIFAR-10AG NewsIMDb ReviewsUCI Adult IncomeUCI German CreditMRI ADCCFD
Metrics
AccuracyEDIDel AUCIns AUCMI AccMI Adv∆Top@128∆EDI
AI summaryThe paper proposes an interpretable federated learning framework, xFedAlign: clients generate private attribution artifacts, and the server aggregates them into a global explanation prior to softly align each party's explanations. On image, text and tabular benchmarks, accuracy is on par with FedAvg, explanation drift is reduced relative to baselines such as Local-XAI, and deletion and insertion AUC improve, with only a few KB of additional communication per round, lower membership inference attack advantage, and stronger resistance to attribution poisoning.

The paper proposes an interpretable federated learning framework, xFedAlign: clients generate private attribution artifacts, and the server aggregates them into a global explanation prior to softly align each party's explanations. On image, text and tabular benchmarks, accuracy is on par with FedAvg, explanation drift is reduced relative to baselines such as Local-XAI, and deletion and insertion AUC improve, with only a few KB of additional communication per round, lower membership inference attack advantage, and stronger resistance to attribution poisoning.

Paper analysis

6 questions; open one to read the full answer

  1. What problem does this paper attempt to solve?

    The authors aim to solve the problem in federated learning of explanations that are both locally faithful and globally consistent while satisfying privacy and bandwidth constraints.

    Core problem: under strict privacy and bandwidth constraints, make federated learning explanations both locally faithful and globally consistent across clients.

    • Problem setting and importance: The authors state that federated learning is being deployed in high-risk domains such as healthcare, finance, and mobile platforms, where auditing and trust calibration require explanations; however, data heterogeneity (non-IID) leads to explanation drift—the same prediction relies on different features across clients (introduction cites Zhang et al., 2024).
    • Shortcomings of existing methods: The authors argue that client-local post-hoc methods preserve local faithfulness but cannot reconcile explanations across clients; server-side aggregation improves comparability but loses fine-grained structure and suppresses minority client behavior; interpretable-by-design models sacrifice accuracy and still drift; sharing gradients, logits, or internal features carries privacy risks and high communication costs.
    • Core observation: The authors argue that existing federated XAI treats explanations as post-hoc artifacts and lacks a mechanism to iteratively coordinate explanation semantics across clients during training; explanation coordination should occur in a low-dimensional, modality-agnostic group probability space rather than in parameter space.
    • Proposed in the paper: xFedAlign, a model-agnostic framework that decouples task optimization in parameter space from explanation coordination in group space: clients distill lightweight surrogate models to produce per-class top-k attribution artifacts, and the server robustly aggregates them into a Global Explanation Prior and softly aligns each client's explanations.
    • Threat model (Appendix C):
      • Attacker goals: membership inference (determining whether a sample is in the training set) and attribution poisoning (manipulating the global explanation under a bounded accuracy loss).
      • Knowledge and capabilities: the membership inference attacker is black box and can see predictions and the broadcast prior, and the strongest variant can also see local explanations of the queried input; the poisoning attacker is a malicious client that is white box with respect to its own surrogate model and local explanations and black box with respect to the aggregated prior and global model.
      • Victim system: a cross-device federated system with K clients plus a central coordinator; the server is honest-but-curious and can only see securely aggregated, noise-added artifacts.
  2. What related research exists?

    Related work falls into three groups: heterogeneous federated optimization, centralized XAI, and federated explanations; the authors state that none of the three coordinates explanation semantics.

    Heterogeneous federated learning optimization

    • FedAvg (McMahan et al., 2017)—establishes the standard paradigm of local optimization plus server aggregation.
    • FedProx, SCAFFOLD, FedAdam, FedBN, FedMOON and similar personalization methods (Li et al., 2020; Karimireddy et al., 2019; Reddi et al., 2020; Li et al., 2021; Smith et al., 2017)—improve stability under heterogeneity through proximal regularization, control variates, adaptive optimizers, normalization, or personalization. The paper notes that these methods coordinate parameters or optimization dynamics, not the semantics of model behavior.
    • Secure aggregation (Bonawitz et al., 2019)—ensures that aggregating model updates does not expose raw client data.

    Centralized XAI

    • Post-hoc attribution: LIME (Ribeiro et al., 2016), SHAP (Lundberg & Lee, 2017), Integrated Gradients (Sundararajan et al., 2017), and gradient-, perturbation-, and visualization-based methods (SmoothGrad, RISE, DeepLIFT, Grad-CAM, etc.).
    • Concept methods and interpretable-by-design models: TCAV (Kim et al., 2018), Concept Bottleneck (Koh et al., 2020), Rudin (2019), etc. The paper notes that most of these methods assume centralized models and static data, do not address the behavior of explanation semantics under heterogeneous sources, and that dense gradients or perturbation mechanisms are costly to transmit and raise privacy concerns.

    Federated explanations

    • Surveys and early explorations: Chaddad et al. (2023) and Li et al. (2023) survey this emerging direction.
    • Federated feature importance and Shapley methods (Wang et al., 2019; Ghorbani & Zou, 2019)—aggregate local attributions without centralizing data; the paper cites empirical studies reporting explanation drift across non-IID clients, especially in medical settings (Rieke et al., 2020; Sheller et al., 2020).
    • Privacy research (Melis et al., 2019; Shokri & Shmatikov, 2015)—points out leakage risks of gradients or logits, relevant to explanation exchange.

    Comparison baselines and benchmarks: the experimental baselines are FedAvg (No-XAI), Local-XAI (Unshared, based on Integrated Gradients), FedAttr-Agg (Wang et al., 2021), and Fed-XAI (Bárcena et al., 2022); the benchmarks are MNIST, CIFAR-10, AG News, IMDb, UCI Adult, and UCI German Credit (the appendix additionally has MRI AD and CCFD).

    Positioning of this paper: the authors state that prior work treats privacy and bandwidth as constraints rather than as first-class parts of the explanation coordination problem; xFedAlign follows the direction of "task learning unchanged, coordinating explanation semantics with compact privacy-aware signals," aligning explanations in a low-dimensional group space via a robustly shared prior.

  3. How does the paper address this problem?

    xFedAlign keeps task optimization in parameter space and coordinates explanations in the group probability space using surrogate models plus a global prior.

    Overall idea: xFedAlign decouples task optimization (parameter space) from explanation coordination (a low-dimensional, modality-agnostic group space G, normalized to the simplex); alignment acts only on explanation distributions and does not change standard federated optimization dynamics.

    Problem setting

    • Cross-device FL: K clients plus a central coordinator; the task model is treated as a black box, and clients do not share raw data, gradients, logits, or internal activations.
    • Explanations are produced and exchanged only in the group space G, normalized to the simplex to ensure cross-client comparability; the goals are local faithfulness, semantic consistency under non-IID, strict privacy, and low communication.

    Clients: adaptive surrogates and local objectives

    • Each client trains a lightweight surrogate gi, using temperature-T distillation with confidence weighting to imitate the task model, with a regularizer promoting sparsity and group structure: Lsur(gi;θ) = Ex∼ Di[KL(pθ(x) | qgi(x))] + λ Ω(gi), i.e., making the surrogate reproduce the task model's class distribution on local data.
    • The attribution operator is path-integrated gradients in the group space starting from a zero baseline, then normalized to the simplex; averaging over the local distribution yields per-class client summaries.
    • Clients keep the top-k groups per class (index–magnitude pairs), apply ℓ2 clipping (bounded sensitivity), quantization (bandwidth), and lightweight Gaussian noise (privacy enhancement), producing artifact Si.

    Server: Global Explanation Prior and alignment

    • The server receives only compact artifacts via secure aggregation, aggregates them with a coordinate-wise median or trimmed mean (RobAgg) and normalizes, obtaining a per-class Global Explanation Prior Π, then broadcasts (θ, Π).
    • Alignment uses Jensen–Shannon divergence (symmetric, bounded [0, log 2], still defined when supports differ). Each round the client objective is the task loss plus α times the surrogate distillation loss plus β times the JSD alignment to Π (Eq. 3).
    • The authors state that because alignment acts only on normalized group distributions, task optimization is unconstrained; robust aggregation plus explanation-space alignment contracts cross-client dispersion (Appendix A gives propositions on dispersion contraction, surrogate faithfulness, (ε, δ)-DP, and convergence).

    Algorithm flow (Algorithm 1)

    1. The server broadcasts (θ(t), Π(t)); clients run one epoch of local task SGD.
    2. Every Rsur rounds, refit the surrogate; compute attributions per class, average, top-k projection, ℓ2 clipping to rclip, quantize to b bits, add Gaussian noise.
    3. Take one JSD alignment gradient step on surrogate parameters; upload model updates and artifacts via secure aggregation.
    4. The server weighted-averages parameters, takes the coordinate-wise median of per-class artifacts and normalizes to obtain the new prior.

    Key hyperparameters (Section 4.4 and Appendix E)

    • 8 clients, 15 rounds, 1 local epoch per round; batch 64, SGD momentum 0.9, lr 0.01, seed 1337.
    • k=128 (MNIST, G=784), ℓ2 clipping radius 5.0, 8-bit quantization, σ=0.1; β linearly warmed up to 0.2 over the first 6 rounds; surrogate updated every R=2 rounds, T=3.0, surrogate lr 0.1, λ=1e-4.
    • Communication is about kC(⌈log2 G⌉+b)/8 bytes, about 2.9KB per round on MNIST (Section 5.4).
  4. What experiments did the paper conduct?

    Main experiments on 6 benchmarks × IID/non-IID, plus privacy, poisoning, ablation, and 128-client scaling experiments on MNIST.

    Experimental setup

    • Evaluated subjects: five schemes—FedAvg (No-XAI), Local-XAI, FedAttr-Agg, Fed-XAI, xFedAlign; same-modality schemes share task architectures (a small CNN for vision, TextCNN for text, a shallow MLP for tabular, Appendix E Table 6).
    • Datasets: MNIST, CIFAR-10, AG News, IMDb, UCI Adult, UCI German Credit, each split into 8 clients; IID uses Dirichlet α=10^6 approximating uniform, non-IID uses α=0.1 label skew; test sets are centralized. RQ2/RQ3 (privacy, poisoning, ablation) use only MNIST.
    • Metrics: Accuracy, EDI (mean JSD between client explanations and the method's respective reference, lower is better), Del AUC (lower is better), Ins AUC (higher is better), MI Acc/MI Adv, ∆Top@128 and ∆EDI.
    • Attack evaluation: membership inference uses a confidence attack with a shadow model plus logistic regression (Appendix G.1); poisoning has malicious clients submit manipulated artifacts, with ρ the fraction of compromised clients (Appendix G.2).
    • Repetitions: all results are mean±std over 5 independent seeds, resampling the client partition each time.

    Main results

    Table 1 (mean±std, 5 runs; EDI/Del AUC lower is better, Accuracy/Ins AUC higher is better):

    Swipe horizontally to read the table

    DatasetPartitionMethodAccuracyEDIDel AUCIns AUC
    MNISTIIDxFedAlign0.986±0.0010.000±0.0000.147±0.0020.958±0.003
    MNISTnon-IIDxFedAlign0.930±0.0040.076±0.0060.170±0.0080.843±0.011
    MNISTnon-IIDLocal-XAI0.927±0.0040.110±0.0060.337±0.0120.737±0.011
    CIFAR-10non-IIDxFedAlign0.460±0.0070.027±0.0030.119±0.0090.327±0.009
    AG Newsnon-IIDxFedAlign0.710±0.0060.166±0.0080.263±0.0100.735±0.011
    IMDbnon-IIDxFedAlign0.702±0.0070.043±0.0080.485±0.0150.568±0.011
    UCI Adultnon-IIDxFedAlign0.799±0.0040.004±0.0010.609±0.0110.772±0.013
    UCI CreditIIDxFedAlign0.735±0.0130.009±0.0020.397±0.0120.628±0.012

    (Per Table 1; FedAvg has no explanation mechanism, so EDI/Del/Ins columns are "–"; rows for Fed-XAI and FedAttr-Agg are omitted.)

    • The authors state that xFedAlign matches FedAvg in accuracy on six benchmarks (MNIST IID 0.986 vs. 0.986), has the lowest EDI in most settings, and improves deletion/insertion AUC.
    • Exceptions: on MNIST IID, Fed-XAI's Del AUC is 0.146±0.001, slightly lower than xFedAlign's 0.147±0.002; on AG News IID, Local-XAI's Ins AUC is 0.794±0.012, higher than xFedAlign's 0.784±0.011; Fed-XAI's accuracy drops substantially under non-IID (MNIST 0.449±0.006).
    • Appendix I: on MRI AD non-IID, xFedAlign Accuracy is 0.7792±0.014 and EDI 0.0018±0.001, higher in accuracy than all baselines; on CCFD, accuracies are close across methods, and xFedAlign's EDI drops to the 10^-3 order (Table 9).

    Privacy and attribution poisoning (Table 2, MNIST)

    • Membership inference: xFedAlign has the lowest MI Acc under both IID and non-IID (0.434±0.010 / 0.463±0.007), and the lowest MI Adv under non-IID (0.004±0.002); under IID, Local-XAI's MI Adv is 0.001±0.001, lower than xFedAlign's 0.004±0.002.
    • Poisoning robustness: xFedAlign has the smallest ∆Top@128 and ∆EDI (IID 0.122±0.009 / 0.003±0.001; non-IID 0.163±0.010 / 0.003±0.001).
    • Authors' interpretation: transmitting only lightly noised top-k artifacts, median aggregation, and JSD alignment together reduce leakage and weaken adversarial influence; Appendix Table 7 shows that at ρ=0.4, xFedAlign's ∆(1−overlap@128) is 0.3828±0.0042, the same as the baselines, and is lower only at intermediate strengths (ρ=0.2, 0.3).

    Ablation (Table 3, MNIST)

    • Alignment weight β: as β rises from 0 to 0.4, EDI drops from 0.0925 to 0.0793, and Del/Ins AUC are essentially unchanged; the β=0 row corresponds to a Local-XAI-style variant without the alignment loss.
    • Noise σ: at σ=0, EDI is 0.1330±0.0091, lower for σ∈[0.05, 0.1], and rises back to 0.0852±0.0012 at σ=0.2; the authors state that light noise acts as implicit regularization.
    • top-k and surrogate epochs: performance is better around k∈{128, 256}; at e=2, 4, EDI rises to 0.0960, 0.1161, based on which the authors choose e=1.

    Other analyses

    • Scaling (Appendix H, Table 8): as clients go from 8 to 128, xFedAlign's EDI stays at 0.0002±0.0000, accuracy drops from 0.9849±0.0012 to 0.8926±0.0143, and at all scales it is not lower than FedAvg.
    • Overhead (Section 5.4, Appendix Table 10): on MNIST IID, GPU training is 39.10 s vs. FedAvg's 32.15 s, and CPU is about 1.7×; about 10 KB extra communication per round; the surrogate has about 10,000 parameters and is not uploaded.
    • Qualitative results (Appendix F, Figures 3, 4): the authors state that xFedAlign's MNIST heatmaps align better with stroke contours, and on UCI Adult importance concentrates on capital gain and education-num; the figures show heatmaps and bar charts for each method, but the converted text does not contain pixel-level details.
  5. What are some points that can be further explored?

    The authors list limitations such as group-space stability and privacy completeness, and plan to extend to regression, federated language models, and adaptive attackers.

    Limitations and future directions noted by the authors

    • Group-space assumption (Section 6): the framework assumes the group space G is meaningful and relatively stable; when groups are ambiguous or feature semantics drift, the prior may need adaptive grouping or drift-aware updates.
    • Alignment does not guarantee other properties (Impact Statement): explanation alignment improves cross-client consistency but does not guarantee fairness, calibration, or causal faithfulness.
    • Privacy assumptions (Impact Statement): the privacy guarantee assumes an honest-but-curious server, secure aggregation, and an artifact-level Gaussian mechanism for DP; Byzantine adversaries beyond the breakdown point of robust aggregation require additional defenses.
    • Sparse artifacts still leak (Appendix C.4): top-k patterns can leak coarse-grained statistics of client populations; when secure aggregation is partially broken, the actual privacy budget is weaker than the theoretical value; sparse artifacts may be susceptible to reconstruction attacks exploiting the structure of the group space.
    • Evaluation environment is controlled (Appendix C.4): membership inference and poisoning are evaluated under a controlled FL setting, standard attribution tools, and moderate noise; large-scale cross-device deployments or richer explanation interfaces may have uncovered leakage channels.
    • Future directions (Section 6): extend to regression tasks, federated language models, and adaptive attackers; strengthen formal privacy and convergence guarantees for coupled task–explanation learning; introduce expert-in-the-loop research to evaluate the practical utility of aligned explanations.

    Scope of experiments

    • The evaluated subjects are five federated explanation schemes sharing three task architectures; benchmarks are 6 standard datasets plus MRI AD and CCFD in the appendix, each with 8 clients and both IID and non-IID partitions (Section 4.1, Appendix I).
    • Privacy, poisoning, and ablation experiments are conducted only on MNIST (Section 4.1); poisoning experiments go up to ρ=0.4, and at ρ=0.4 xFedAlign's ∆(1−overlap@128) is the same as the baselines (Table 7).
    • All results are mean±std over 5 seeds (Section 5); scaling experiments cover up to 128 clients (Table 8).
    • The membership inference attack is a confidence-based shadow attack (Appendix G.1); the paper does not report results for membership inference variants other than this attack.
    • The paper discusses parameter-space regularizers such as FedProx and SCAFFOLD, stating that the two are complementary and can be stacked, but does not report experimental numbers after stacking (Section 3.3).
  6. Summarize the main content of the paper

    xFedAlign coordinates federated explanations in group space; the authors say it maintains accuracy, reduces drift, and enhances privacy and poisoning resistance.

    A paper on a federated interpretability method: the authors propose a model-agnostic framework, xFedAlign, addressing explanation drift in non-IID federated learning, loss of structure in aggregation, and the privacy and bandwidth costs of sharing dense attributions.

    • Problem: local post-hoc explanations drift across heterogeneous clients, global aggregation suppresses minority-client behavior, and interpretable-by-design models sacrifice accuracy; the authors argue there is a lack of mechanisms to coordinate explanation semantics during training.
    • Method: task optimization stays in parameter space; each client trains a lightweight surrogate via temperature distillation, producing per-class top-k attributions in a normalized group space, which after ℓ2 clipping, 8-bit quantization, and Gaussian noise are aggregated by the server via coordinate-wise median into a Global Explanation Prior, then a JSD alignment loss softly constrains each client's explanations; Appendix A gives propositions on dispersion contraction, surrogate faithfulness, (ε, δ)-DP, and convergence.
    • Main results (Table 1, 5 seeds): accuracy matches FedAvg on six benchmarks (MNIST IID both 0.986±0.001), EDI is lowest in most cases (MNIST IID 0.000±0.000; UCI Adult non-IID 0.004±0.001); the exception is AG News IID, where Ins AUC is lower than Local-XAI.
    • Privacy and robustness (Table 2, MNIST): MI Acc is lowest under both partitions (0.434 / 0.463), and ∆Top@128 and ∆EDI are smallest under poisoning; however, at ρ=0.4, ∆(1−overlap@128) is the same as the baselines (Table 7).
    • Scaling and overhead: with 128 clients, EDI stays at 0.0002 (Table 8); on MNIST, about 10 KB extra communication per round, and GPU training time about 1.2× FedAvg (Table 10).
    • Authors' conclusion: separating task space from explanation space is a viable design principle for trustworthy federated learning; they also acknowledge limitations such as group-space stability and privacy completeness, and plan to extend to regression, federated language models, and adaptive attackers.

Related papers

Both use: AG News
Show 3 more papers
Read originalICML 2026