In federated learning, many devices train a shared model together. Each device sends an update to a central server, and the server combines them. To survive devices that are broken or malicious, the field uses "robust aggregators," rules that throw out or down-weight updates that look unlike the rest. These rules rest on the majority-benign assumption: they work as long as honest devices outnumber attackers. We argue that this single threshold describes failure incorrectly. Honest devices with different local data do not agree on one update. They form several internally coherent groups, each a legitimate route to a good shared model. A tightly coordinated attacker bloc then only needs to outnumber the largest honest group, not all honest devices together. On this view, breakdown is governed by plurality rather than majority, and it can be reached by several different combinations of conditions.
To test this, we borrow fuzzy-set qualitative comparative analysis (fsQCA), a social-science method that finds which combinations of conditions are sufficient for an outcome. It allows several distinct combinations to produce the same outcome ("equifinality"). It also allows the recipes for failure to differ from the mirror image of the recipes for success. We apply it to a large, controlled grid of federated training runs across the main defense families. The deliverable is a tested map of the recipes that break each defense, together with one sharp, falsifiable prediction: at a fixed global attacker share, peer-consensus defenses break when attackers outnumber the largest honest mode. Validation-data defenses break through a different recipe, which is poor coverage of honest modes in the server's trial data.
Core Claim
The claim has two layers. The substantive layer concerns federated learning. Let $f$ be the number of coordinated attackers sampled in a round and $m_{\max}$ the number of sampled honest clients in the largest honest mode, meaning the largest group of honest clients whose updates are mutually coherent because their local data are similar. We hypothesize that selection-based aggregators and clustering filters break when the plurality margin
$$\rho = \frac{f}{m_{\max}}$$
crosses roughly one, even when $f$ is well below $n/2$, where $n$ is the number of sampled clients. The form follows from how these rules choose. Krum, from Machine Learning with Adversaries: Byzantine Tolerant Gradient Descent, scores each update by its distances to its nearest neighbors and keeps the tightest one. The tightest neighborhood is therefore whichever coherent bloc is largest. Krum's guarantee $2f+2<n$ was proved under a single true gradient direction, which is exactly the assumption we relax. For coordinate-wise rules such as median and trimmed mean, the prediction is less sharp, because honest modes on opposite sides of a coordinate can partly protect the median. That uncertainty is one reason a configurational analysis is needed rather than a single threshold.
The methodological layer concerns fsQCA itself. Its known weaknesses are small samples, empty truth-table rows, and arbitrary calibration. These largely disappear when the researcher controls a simulator and can fill every row. Federated robustness is therefore a setting where the method's claims about equifinality and asymmetry can be checked against held-out ground truth, which observational uses of the method cannot do.
Relation to Prior Work
The federated learning papers motivate the gap by showing that "majority" is already leaky. The Breakdown Point of Byzantine Robust Federated Learning collects evidence that median-style rules fail at 50% attackers. It also records that when attackers hold the majority, distance-based filters invert: they purge honest heterogeneous clients and keep the poisoned center. The same line of work reports that random client sampling can give a 10% global attacker population a transient local majority in some rounds. Both findings treat the honest population as one cloud. We ask what happens when that cloud is several clouds.
Shielding Federated Learning: Robust Aggregation with Adaptive Client Selection (MAB-RFL) acknowledges that non-IID honest clients point in different directions and holds up to 48% attackers. It still models diversity as scatter around one consensus. FLDetector avoids server data by flagging clients whose updates stray from their own predicted trajectory. That design would misread honest clients that legitimately change modes.
The validation-data defenses constrain the design, because they are the documented way past 50%. Zeno ranks updates by loss reduction on a trusted server batch and needs only one honest worker. Bant weights updates by their effect on a server trial function and reports that its automatic variant diverges when the trial set is very small. FedGame reverse-engineers backdoor triggers with auxiliary data. Practical Differentially Private and Byzantine-resilient Federated Learning uses as few as two auxiliary samples per class. All four score updates against a single objective. None of them asks what happens when the trial data represent some honest modes and not others. We predict that this coverage condition is a second, distinct failure recipe that operates even at low attacker share.
The fsQCA and equifinality papers supply the method. Examples include Institutional Configuration Study of UrbanGreen Economic Efficiency – AnalysisBased on fsQCA and NCA, which pairs fsQCA with Necessary Condition Analysis (NCA, a test for conditions without which the outcome cannot occur). Synergistic Innovation Pathways in Aviation Complex Product Ecosystems shows that the configurations producing high outcomes are not mirrors of those producing low ones. We reuse their calibration, truth-table minimization, and consistency and coverage metrics unchanged. Our contribution is to apply them where the ground truth can be controlled.
The research file contains no work applying set-theoretic analysis to federated robustness. We also deliberately do not pursue the most obvious combination, using fsQCA as a data-free detector that breaks the 50% barrier. When a coordinated attacker majority is structurally indistinguishable from honest clients, no peer-relative rule can tell which bloc is honest. A Boolean relabeling of cluster features would not change that.
Research Program
We build a simulation grid over six conditions. We cap the number at six so that the 64-row truth table can be fully populated. The conditions are:
- Global attacker share, from 10% to 70%.
- Participation rate, which controls how much the per-round attacker share varies.
- Largest-honest-mode share, set by assigning honest clients to label-skewed groups of controlled sizes, or to writer groups in FEMNIST.
- Attacker coordination, ranging from independent noise to identical Sybil updates.
- Trial-data coverage of honest modes, the fraction of modes represented in the server batch, which is zero for data-free defenses.
- Attack family, either untargeted sign-flip and ALIE-style attacks or scaling backdoors.
Each case is one run of one defense under one configuration and seed. The defense families are:
- Peer consensus: Krum, Multi-Krum, coordinate-wise median, and trimmed mean.
- Clustering and selection: MAB-RFL.
- Temporal consistency: FLDetector.
- Validation-based scoring: Zeno and Bant.
- An unprotected reference: plain averaging.
Calibration uses substantive anchors rather than sample quantiles. For clean-task outcomes, "broken" has full membership at within five points of chance accuracy, crossover at the midpoint between chance and the no-attack accuracy of the same configuration, and full non-membership within two points of no-attack accuracy. For backdoors, attack success rate plays the same role. We run fsQCA separately for the presence of breakdown and for robust survival, and run NCA for necessity. We then check sensitivity by shifting every anchor by ±10% and reporting which recipes survive. Where applicable, attacks and defenses will be implemented on microsoft/robustlearn, with the remaining defenses reimplemented from their papers under a shared training loop.
Baselines
The naive reference is the majority-benign rule itself. It predicts breakdown if and only if the per-round attacker share reaches the defense's stated tolerance: 50% for median, the $2f+2<n$ bound for Krum, and effectively never for Zeno and Bant. This is the explanation the field currently uses. It must be beaten on held-out configurations, or the whole claim fails.
The strongest competing explanation is a shallow decision tree or rule list, specifically a depth-limited classification tree. It is fit to the same case table, with the same six raw conditions as inputs, the same train and test split of configuration cells, and an equal hyperparameter search budget. A tree also produces conjunctive rules, so it is the fair test of whether fsQCA adds anything beyond interpretable rule learning. A second statistical competitor is logistic regression with all pairwise interactions, which represents the "net effects plus interactions" view that fsQCA claims to improve on.
The key ablation reruns every method with the largest-honest-mode condition replaced by a conventional non-IID knob, the Dirichlet concentration of label skew. The Dirichlet knob controls heterogeneity without controlling mode structure. If prediction degrades only when mode structure is removed, the plurality variable carries the result.
We also anchor defense performance to reported numbers so that our reimplementations are credible. On CIFAR-10 with 40% attackers, MAB-RFL reports about 73% accuracy while competing defenses fall to 10%. FLDetector reports restoring CIFAR-10 accuracy from 27.8% under untargeted attack to 65.9%, against a 65.8% no-attack baseline. FedGame reports CIFAR-10 attack success rate below 13% under a scaling attack. The differentially private defense reports about 80% Fashion-MNIST accuracy with 40% attackers. There is no reported result on Lacuna for Zeno's or Bant's CIFAR-10 accuracy, so for those two we reproduce only their qualitative claim of surviving a majority of attackers in the IID case.
Beating the baselines means two things. The fsQCA solution, or a tree using the plurality variable, must cut held-out breakdown-prediction error well below the majority rule. The fsQCA solution must also be at least as accurate as the tree while recovering the separately predicted failure recipes with consistency above 0.8.
Why not just…?
Why not just run a factorial sweep and fit a decision tree? This is the real threat, and we treat it as a co-equal analysis rather than a straw man. A tree fit to failure is forced to describe success as the complement of failure. fsQCA fits the two outcomes separately and scores each recipe's consistency and coverage, which is the right tool if survival and breakdown have different causes. The held-out comparison settles which is better. If the tree matches fsQCA, we will report that the contribution is the plurality variable and the coverage variable, not the set-theoretic machinery. That outcome would still overturn the majority-benign threshold.
Why not just sweep the attacker share, as every defense paper does? One-dimensional sweeps are the source of the majority threshold. The breakdown-point line of work already shows that per-round sampling changes the effective share. A sweep over attacker share with a fixed heterogeneity recipe cannot reveal that a 30% attacker bloc beats a population split into four honest modes. Our fixed-share experiment below targets exactly that case.
Why not just use Zeno or Bant and declare the problem solved? They do survive attacker majorities in the reported settings. They also score every update against one server objective, and Bant is documented to diverge with small trial sets. Our coverage condition tests whether honest modes that are missing from the trial data get purged even at 10% attackers. If it does not matter, these defenses are exonerated on this axis, and our map will say so.
Proposed Experiments and Decisive Checks
The first decisive check is the fixed-share experiment on FEMNIST and label-skewed CIFAR-10 with 50 to 100 sampled clients per round. We hold the global attacker share at 30% with coordinated Sybil updates and vary the largest honest mode from 70% of honest clients down to 15%. The plurality hypothesis predicts that Krum and MAB-RFL-style clustering degrade sharply near $\rho \approx 1$, while plain honest heterogeneity at matched Dirichlet skew does not produce the same drop. If accuracy stays flat across $\rho$, the plurality claim is false for that defense.
The second check is the coverage experiment for Zeno and Bant on CIFAR-10 and Fashion-MNIST. We vary the fraction of honest modes represented in the trial batch at attacker shares of 10%, 30%, and 60%. The prediction is a failure recipe of low coverage combined with a minority honest mode. It should appear as a separate fsQCA path from the peer-consensus path, which is the equifinality claim.
The third check is transfer. We fit recipes on CIFAR-10 and FEMNIST runs and predict breakdown on Fashion-MNIST configurations never seen during analysis, scoring all baselines by balanced accuracy.
The fourth check is asymmetry. We test whether the survival solution contains conditions that are absent from the negation of the breakdown solution. Coordination is a likely candidate, since it may matter for survival only when trial coverage is high.
Negative outcomes on any of these checks are informative, and the study is designed so that each can come out against us.
References
- Machine Learning with Adversaries: Byzantine Tolerant Gradient Descent
- Resilient Primal-Dual Optimization Algorithms for Distributed Resource Allocation
- Shielding Federated Learning: Robust Aggregation with Adaptive Client Selection
- FLDetector: Defending Federated Learning Against Model Poisoning Attacks via Detecting Malicious Clients
- Zeno: Byzantine-suspicious stochastic gradient descent
- Bant: Byzantine Antidote via Trial Function and Trust Scores
- FedGame: A Game-Theoretic Defense against Backdoor Attacks in Federated Learning
- Practical Differentially Private and Byzantine-resilient Federated Learning
- Institutional Configuration Study of UrbanGreen Economic Efficiency – AnalysisBased on fsQCA and NCA
- Synergistic Innovation Pathways in Aviation Complex Product Ecosystems: Enabling Sustainability Through Resource Efficiency and Systemic Collaboration