Explainable Federated Learning via Global-Local Attribution Alignment
Organizations: Department of Computer Science, Virginia Tech, Falls Church, VA, USA · DEVCOM Army Research Laboratory, Adelphi, MD, USA
Abstract
Federated learning enables on-device training without centralizing data, yet existing systems still struggle to provide explanations that are both locally faithful and globally consistent under strict privacy and bandwidth constraints. Prior approaches either keep explanations siloed across clients, transmit heavy or sensitive artifacts, or replace expressive task models with interpretable surrogates that sacrifice accuracy. We propose xFedAlign, a model-agnostic framework that decouples task optimization in parameter space from explanation coordination in a compact group space. Each client distills a lightweight surrogate to produce private, per-class top-k attribution artifacts, which are robustly aggregated by the server into a Global Explanation Prior that softly aligns client explanations without constraining task learning. Across image, text, and tabular benchmarks with IID and non-IID partitions, xFedAlign matches FedAvg accuracy while consistently reducing explanation drift and improving deletion and insertion AUC relative to Local-XAI, FedAttr-Agg, and Fed-XAI, with only a few kilobytes of additional communication per round. Privacy and robustness evaluations further demonstrate reduced membership inference advantage and increased resistance to attribution poisoning, enabling consistent and trustworthy explanations in federated learning.
Figures & tables
| Type | Dataset | Approach | IID | Non-IID | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Accuracy | EDI | Del AUC | Ins AUC | Accuracy | EDI | Del AUC | Ins AUC | |||
| Image | MNIST | FedAvg | 0.986 0.001 | – | – | – | 0.930 0.004 | – | – | – |
| Local-XAI | 0.982 0.002 | 0.020 0.001 | 0.189 0.009 | 0.956 0.003 | 0.927 0.004 | 0.110 0.006 | 0.337 0.012 | 0.737 0.011 | ||
| FedAttr-Agg | 0.983 0.002 | 0.077 0.002 | 0.174 0.004 | 0.939 0.004 | 0.930 0.004 | 0.159 0.007 | 0.288 0.010 | 0.747 0.010 | ||
| Fed-XAI | 0.904 0.001 | 0.001 0.000 | 0.146 0.001 | 0.750 0.002 | 0.449 0.006 | 0.300 0.009 | 0.126 0.003 | 0.425 0.009 | ||
| xFedAlign | 0.986 0.001 | 0.000 0.000 | 0.147 0.002 | 0.958 0.003 | 0.930 0.004 | 0.076 0.006 | 0.170 0.008 | 0.843 0.011 | ||
| Method | MI Acc (↓) | MI Adv (↓) | Top@128 (↓) | EDI (↓) |
|---|---|---|---|---|
| IID | ||||
| FedAvg | 0.460 0.007 | 0.005 0.002 | 0.231 0.012 | 0.005 0.001 |
| Local-XAI | 0.460 0.006 | 0.001 0.001 | 0.231 0.011 | 0.005 0.001 |
| FedAttr-Agg | 0.460 0.008 | 0.005 0.002 | 0.231 0.010 | 0.005 0.001 |
| Fed-XAI | 0.446 0.009 | 0.007 0.003 | 0.233 0.013 | 0.007 0.002 |
| xFedAlign | 0.434 0.010 | 0.004 0.002 | 0.122 0.009 | 0.003 0.001 |
| Ablation | Value | EDI | DelAUC | InsAUC |
|---|---|---|---|---|
| (alignment) | 0.0 | 0.0925\textsubscript{\scriptsize\pm0.0006} | 0.1715\textsubscript{\scriptsize\pm0.0067} | 0.9006\textsubscript{\scriptsize\pm0.0230} |
| 0.05 | 0.0903\textsubscript{\scriptsize\pm0.0012} | 0.1696\textsubscript{\scriptsize\pm0.0042} | 0.8997\textsubscript{\scriptsize\pm0.0228} | |
| 0.1 | 0.0876\textsubscript{\scriptsize\pm0.0012} | 0.1699\textsubscript{\scriptsize\pm0.0045} | 0.8999\textsubscript{\scriptsize\pm0.0227} | |
| 0.2 | 0.0841\textsubscript{\scriptsize\pm0.0007} | 0.1691\textsubscript{\scriptsize\pm0.0056} | 0.9001\textsubscript{\scriptsize\pm0.0209} | |
| 0.4 | \mathbf{0.0793\textsubscript{\scriptsize\pm0.0008}} | 0.1690\textsubscript{\scriptsize\pm0.0035} | 0.8996\textsubscript{\scriptsize\pm0.0230} | |
| (DP noise) | 0.0 | 0.1330\textsubscript{\scriptsize\pm0.0091} | 0.1698\textsubscript{\scriptsize\pm0.0034} | \mathbf{0.9009\textsubscript{\scriptsize\pm0.0225}} |
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Symbol | Meaning |
|---|---|
| Number of federated clients | |
| Set of clients participating in round | |
| Number of classes | |
| Local dataset on client | |
| Task model with parameters | |
| Surrogate model on client |
| Modality | Dataset | Input type | Classes | Train | Val / Test |
|---|---|---|---|---|---|
| Image | MNIST | grayscale | 10 | 54,000 | 6,000 / 10,000 |
| CIFAR-10 | RGB | 10 | 45,000 | 5,000 / 10,000 | |
| Text | AG News | Token sequences | 4 | 108,000 | 12,000 / 7,600 |
| IMDb Reviews | Token sequences | 2 | 22,500 | 2,500 / 25,000 | |
| Tabular | UCI Adult Income | Continuous + categorical | 2 | 29,305 | 3,256 / 16,281 |
| UCI German Credit | Continuous + categorical | 2 | 630 | 70 / 300 |
| Modality | Dataset | Architecture | Parameters |
|---|---|---|---|
| Image | MNIST | Input ; + BN + ReLU; + BN + ReLU; max pool; dropout ; FC + ReLU; dropout ; FC | M |
| Image | CIFAR–10 | Input ; + BN + ReLU; + BN + ReLU; max pool; + BN + ReLU; + BN + ReLU; max pool; dropout ; FC + ReLU; FC | M |
| Text | AG News | Tokenized sequence ; embedding with ; 1D convs with kernel widths , filters per ; global max pool over time for each ; concatenation ; dropout ; FC | M |
| Text | IMDb Reviews | Same encoder family as AG News with filters per kernel, ; dropout ; FC | M |
| Tabular | UCI Adult Income | Preprocessed input with standardized continuous features and one-hot categorical features ( ); FC + ReLU; dropout ; FC + ReLU; FC | |
| Tabular | UCI German Credit | Preprocessed input with ; FC + ReLU; dropout ; FC + ReLU; FC |
| Metric | FedAvg | Local-XAI | FedAttr-Agg | Fed-XAI | xFedAlign | |
|---|---|---|---|---|---|---|
| 0.00 | 0.0000\textsubscript{\scriptsize\pm0.0000} | 0.0000\textsubscript{\scriptsize\pm0.0000} | 0.0000\textsubscript{\scriptsize\pm0.0000} | 0.0000\textsubscript{\scriptsize\pm0.0000} | 0.0000\textsubscript{\scriptsize\pm0.0000} | |
| 0.10 | 0.0078\textsubscript{\scriptsize\pm0.0005} | 0.0078\textsubscript{\scriptsize\pm0.0005} | 0.0078\textsubscript{\scriptsize\pm0.0005} | 0.0156\textsubscript{\scriptsize\pm0.0007} | 0.0000\textsubscript{\scriptsize\pm0.0000} | |
| 0.20 | 0.3828\textsubscript{\scriptsize\pm0.0041} | 0.3828\textsubscript{\scriptsize\pm0.0040} | 0.3828\textsubscript{\scriptsize\pm0.0042} | 0.3828\textsubscript{\scriptsize\pm0.0043} | 0.0156\textsubscript{\scriptsize\pm0.0008} | |
| 0.30 | 0.3828\textsubscript{\scriptsize\pm0.0043} | 0.3828\textsubscript{\scriptsize\pm0.0042} | 0.3828\textsubscript{\scriptsize\pm0.0041} | 0.3828\textsubscript{\scriptsize\pm0.0044} | 0.2109\textsubscript{\scriptsize\pm0.0035} | |
| 0.40 | 0.3828\textsubscript{\scriptsize\pm0.0040} | 0.3828\textsubscript{\scriptsize\pm0.0040} | 0.3828\textsubscript{\scriptsize\pm0.0042} | 0.3828\textsubscript{\scriptsize\pm0.0041} | 0.3828\textsubscript{\scriptsize\pm0.0042} | |
| 0.00 | 0.0000\textsubscript{\scriptsize\pm0.0000} | 0.0000\textsubscript{\scriptsize\pm0.0000} | 0.0000\textsubscript{\scriptsize\pm0.0000} | 0.0000\textsubscript{\scriptsize\pm0.0000} | 0.0000\textsubscript{\scriptsize\pm0.0000} |
| Method | Acc | EDI | Del. AUC | Ins. AUC | |
|---|---|---|---|---|---|
| 8 | FedAvg | – | – | – | |
| Local-XAI | |||||
| FedAttr-Agg | |||||
| Fed-XAI | |||||
| xFedAlign | |||||
| 16 | FedAvg | – | – | – |
| Dataset | Split | Approach | Accuracy | EDI | Del AUC | Ins AUC |
|---|---|---|---|---|---|---|
| MRI AD | IID | FedAvg | 0.7208\textsubscript{\scriptsize\pm0.015} | – | – | – |
| Local-XAI | 0.7479\textsubscript{\scriptsize\pm0.014} | 0.0170\textsubscript{\scriptsize\pm0.002} | 0.4190\textsubscript{\scriptsize\pm0.018} | 0.5740\textsubscript{\scriptsize\pm0.020} | ||
| FedAttr-Agg | 0.7156\textsubscript{\scriptsize\pm0.016} | 0.0984\textsubscript{\scriptsize\pm0.006} | 0.3610\textsubscript{\scriptsize\pm0.017} | 0.5730\textsubscript{\scriptsize\pm0.018} | ||
| Fed-XAI | 0.5000\textsubscript{\scriptsize\pm0.020} | 0.1329\textsubscript{\scriptsize\pm0.007} | 0.4000\textsubscript{\scriptsize\pm0.016} | 0.4370\textsubscript{\scriptsize\pm0.019} | ||
| xFedAlign | 0.7115\textsubscript{\scriptsize\pm0.015} | 0.0018\textsubscript{\scriptsize\pm0.001} | 0.3370\textsubscript{\scriptsize\pm0.015} | 0.6260\textsubscript{\scriptsize\pm0.021} | ||
| Non-IID | FedAvg | 0.4635\textsubscript{\scriptsize\pm0.018} | – | – | – |
| CPU time (s) | GPU time (s) | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | Train | Expl | Total | Train | Expl | Total | Task params | Surrogate params (on-device) | Comm/Rnd (KB) | |
| FedAvg | 213.04 | 0.00 | 213.04 | 32.15 | 0.00 | 32.15 | 406,922 | – | 1589.54 | |
| Local-XAI | 213.38 | 0.13 | 213.50 | 32.47 | 0.02 | 32.49 | 406,922 | – | 1589.54 | |
| FedAttr-Agg | 215.93 | 0.12 | 216.05 | 33.01 | 0.02 | 33.03 | 406,922 | – | 1591.45 | |
| Fed-XAI | 101.60 | 0.00 | 101.60 | 14.85 | 0.00 | 14.85 | 7,850 | – | 30.66 | |
| xFedAlign | 355.18 | 0.00 | 355.18 | 39.10 | 0.00 | 39.10 | 406,922 | 10,000 | 1599.54 | |
Federated Learning
Federated learning trains shared models across data holders without requiring them to pool their raw data.
See all 1,027 papers