SMASH 2026 · Paper #16 · Result card

Tracing Bias Propagation Across Subgroups in Federated Learning

Tirathraj Ramburn, Afaf Taïk, Farnoosh Ghadiri · Read the paper

Mislabelling 30% of A1-B3's positives hurts A1-B4 most, 17.1% more than A1-B3 itself. The closed-form prediction matches the retrained ranking at Spearman 0.980.

source error 30% of A1-B3 positives → 0 at client 0clients 4heterogeneity h 0.0β (1,Z,R,W1,W2) -1,0.85,0.90,-0.70,0.55A→R shift 0.25seeds 20analysis IID (8 groups, recipients averaged)

Rank agreement (Spearman)

0.980

95% CI [0.974, 0.986] over 20 seeds

Size agreement (slope c)

1.002

1.0 = measured as large as predicted; CI [0.982, 1.023]

Most affected recipient

A1-B4

17% more than target A1-B3 (rank 2)

Labels changed

45.5

0.379% of 12,000 federation labels; every recipient mean decreased

From the abstract

We show that the impact of false-negative errors on recipient client–subgroup pairs can be predicted from the global model's curvature together with source and recipient-group feature summaries, without retraining a model for each possible error. Notably, a non-target subgroup can experience the largest degradation.

Evidence

Can we predict who gets hurt without retraining?

Audit without retraining: every possible source error

Linearity in ρ, and the direction of propagation

Paper vs. your run

Metric	Paper	Your run
Spearman (IID, 30%)	A1-B3: 0.975 [0.961, 0.987] · A0-B2: 0.968 [0.955, 0.980]	0.980 [0.974, 0.986]	target A1-B3
Slope c	A1-B3: 1.014 · A0-B2: 1.016 (h=1: 1.009 / 1.007)	1.002
Most affected vs target	A1-B3 → A1-B4 hit 18.2% more · A0-B2 → A1-B1 hit 34.9% more (target ranks 4th)	A1-B4, 17.1% more (target rank 2)
Labels changed (30%)	A1-B3: 45.25 (0.377%) · A0-B2: 27.20 (0.227%)	45.50 (0.379%)
Sources whose biggest victim is another group	— (new: audit matrix)	5 of 8	predicted, no retraining

Try next

A1-B1 as the mislabelled group — The audit predicts A1-B1 as its own top victim, unlike A1-B3, so this tests whether same-group damage dominates.
```
{"target_group": "A1-B1"}
```
Heterogeneous clients at h = 1 — Splits recipients by client and tests the paper's heterogeneous setting, where it reports Spearman 0.974 for A1-B3.
```
{"heterogeneity": 1}
```
No association between A and R — With r_shift 0, A0 and A1 recipients should look more alike along the propagation vector.
```
{"r_shift": 0}
```

Read this before sharing

Synthetic data from the paper's generator.
Logistic regression only, as in the paper.