Who Pays for Byzantine Robustness? Validity, Data-Share Dictatorships, and Fairness When Only a Few Hospitals Collaborate

Adnan MalkaniarXiv 2026 · arXiv

In the world of Federated Learning (FL), "Byzantine robustness" is the holy grail of security. It promises that a central model can still learn effectively even if some participating nodes (clients) send malicious or corrupted updates. However, most research in this area assumes a world of dozens or hundreds of equal-sized clients.

In a new paper titled "Who Pays for Byzantine Robustness?", Adnan Malkani from Sapienza University of Rome audits these assumptions against the reality of clinical FL. In medical collaborations, you don't have thousands of phones; you have three to six hospitals with wildly different patient volumes. This "small and skewed" reality changes everything.

TL;DR

Standard Byzantine-robust defenses for federated learning often fail or become "dictatorships" in small medical networks where one hospital holds most of the data. The paper reveals that "robustness" often comes at the cost of fairness, as these algorithms tend to penalize hospitals with unique patient populations rather than just malicious actors.

What Problem Does This Paper Address?

Federated learning allows hospitals to train a shared model without sharing raw patient data. But what if one hospital’s server is hacked, or its imaging equipment is miscalibrated? These are Byzantine failures—updates that can be arbitrarily wrong and can easily "poison" a standard model.

To fix this, researchers use Robust Aggregation Rules (AGRs) like Krum, Bulyan, or Geometric Median. These rules try to find and ignore "outlier" updates. The problem is that these rules were designed for large, balanced datasets. In real-world clinical networks:

  1. The "Small $n$" Problem: Many rules require a minimum number of participants to work (e.g., $n \ge 2f+3$). In a 4-hospital network, some rules are mathematically undefined.
  2. The Skew Problem: One hospital might have 50% of the data while another has 2%. Standard defenses often ignore these size differences or, conversely, give the big hospital too much power.

Methods

Malkani conducted an exhaustive empirical audit of eight standard AGRs (including FedAvg, Median, Krum, and Centered Clipping) across both unweighted and sample-size-weighted variants.

The study's most novel methodological contribution is the "exact-twin leave-out control." To understand if a drop in accuracy is caused by a malicious attack or simply by the loss of a hospital's good data, the author compared attacked runs against "control" runs where the attacker was simply absent from the training. This allowed for a precise decomposition of performance loss into "Honest Data Loss" vs. "True Attack Damage."

Datasets

The audit used three realistic medical datasets from the FLamby benchmark:

  • Fed-Heart-Disease: 4 hospitals (skewed from 6% to 41% data share).
  • Fed-TCGA-BRCA: 6 regions (breast cancer genomics).
  • Fed-ISIC2019: 6 dermatology centers (highly skewed: one center holds 53.4% of the data).

Key Insights and Ideas

  • Data-Share Dictatorships: When using weighted robust rules (like weighted Median), if one hospital holds >50% of the data, it becomes a "dictator." The algorithm effectively ignores everyone else to stay "robust" to the smaller players. If this giant hospital turns malicious, the defense collapses instantly.
  • The Fairness Tax: Unweighted robust rules are "fair" in terms of voting power, but they consistently penalize whichever institution is the most "atypical." In medical data, being an outlier often means you have a unique, valuable patient population (e.g., a specific demographic). Robust rules treat these unique hospitals as potential attackers.
  • Illusion of Failure: When a large hospital attacks and accuracy plunges, we usually blame the "poison." Malkani shows that for unweighted rules, 70-80% of that drop is actually just the loss of the hospital's honest data, not the damage from the attack itself.
Fig. 1. Status of every rule variant in each federation (n, f): invalid by client count, reducing to the mean or median, or flagged by data share as a dictator (the clean output is the largest client’s update) or as breaking when the heaviest f clients attack.

Key Results

  • Validity Breakdown: Bulyan is mathematically invalid for all 4- and 6-hospital networks tested. Krum and Multi-Krum fail in 4-hospital setups.
  • Attack Damage: On the ISIC dataset, when the largest center attacks, unweighted robust rules limited "true attack damage" to just 2.3–3.8 points. However, weighted versions of the same rules suffered massive damage (up to 25.1 points) because they gave the attacker too much influence.
  • The "Winner": Centered Clipping (CC) emerged as a top performer. It incurred almost zero "clean cost" (accuracy loss when no one is attacking) while providing strong defense, especially in its unweighted form.
Figure 2: Relative influence of clients. Note how weighted rules (w) often zero out the influence of smaller hospitals like Switzerland (6% share).

Limitations and Future Work

The study focused on linear and logistic models to allow for the massive scale of 54,000+ training runs. While these are common in clinical research, future work is needed to see if these findings hold for full fine-tuning of large-scale deep learning models or Large Language Models (LLMs).

Additionally, the study assumes hospitals report their true sample sizes. A sophisticated Byzantine attacker might lie about their data volume to further manipulate weighted aggregation rules—a "weight-inflation" attack that remains an open research question.

For those deploying FL in healthcare, the takeaway is clear: don't just "turn on" a robust aggregator. Audit your data distribution first, and consider keeping sample weights out of the filtering step to avoid creating a data-share dictatorship.

Summary updated

Abstract

Clinical federated learning (FL) collaborations typically connect only three to six hospitals, yet Byzantine-robust aggregation rules are designed and benchmarked on synthetic partitions across dozens of equal-sized clients. We conduct an empirical audit of eight standard aggregation rules: FedAvg, coordinate-wise median, trimmed mean, Krum, Multi-Krum, Bulyan, geometric median (GeoMed), and Centered Clipping (CC). We evaluate both unweighted and sample-size-weighted variants across three cross-silo medical datasets from the FLamby benchmark: Fed-Heart-Disease ($n = 4$ hospitals), Fed-TCGA-BRCA ($n = 6$ regions), and Fed-ISIC2019 ($n = 6$ centres). We execute 54,214 training runs across enumerated attacker coalitions ($f \in \{1, 2\}$) paired with exact-twin leave-out controls that decompose performance drops into honest data loss versus adversarial attack damage. We report three findings. First, client-count validity checks fail to capture breakdown in skewed medical consortia: while $n = 6, f = 1$ satisfies every count condition except Bulyan's, assigning sample-size weights turns median and GeoMed into a clean-model dictatorship whenever a single hospital holds more than half the training data, and causes weighted median, GeoMed, and CC to collapse when that dominant centre attacks. Second, exact-twin leave-out controls reveal that when the largest hospital attacks, the apparent failure of count-based robust rules is predominantly caused by the loss of honest training data (7.5 to 10.4 points on ISIC) rather than adversarial perturbation (2.3 to 3.8 points on ISIC). Conversely, weight-dependent rules suffer large true attack damage (16.1 to 25.1 points on ISIC). Third, unweighted robust rules consistently discount whichever institution generates the most atypical parameter updates, penalizing the smallest hospital on Heart but penalizing larger institutions on TCGA and ISIC.
BibTeX
@misc{malkani2026who,
  author = {Adnan Malkani},
  title = {{Who Pays for Byzantine Robustness? Validity, Data-Share Dictatorships, and Fairness When Only a Few Hospitals Collaborate}},
  year = {2026},
  eprint = {2611.22904},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22904},
}
Figure