Organizational research keeps finding the same pattern. When people are rewarded for fitting in and punished for standing out, they stop offering their most distinctive ideas. They pull back toward what the institution already believes, and the institution never learns what it was missing. We propose that a federated learning system can produce the same failure through its own rules. Federated learning means many parties jointly train one model by sending model updates instead of raw data. Once its clients are paid or penalized according to whether the server accepts their updates, a client whose data is unusual, but valuable, has a rational reason to shrink its update toward the consensus. The global model then quietly loses the knowledge that only that client could supply.
Two literatures have never been joined here. Federated learning has defenses that discard unusual updates as possible attacks, and it has incentive schemes that reward or penalize client contributions. Yet the field evaluates defenses on clients that submit honestly, and incentives on clients whose only strategic choice is how much effort to report. We ask what happens when the two are combined and clients can choose how "different" to look. Our main claim is that rewards tied to acceptance or to closeness to consensus produce a measurable "conformity tax" on minority-data clients. We further claim that rewards tied to the update's measured benefit to other clients remove most of it, and can do so without the server holding any validation data.
Where the idea comes from, and what it is not
The organizational studies supply the hypothesis, not the evidence. Liu et al. study "creative deviance," meaning employees who keep working in secret on ideas management has rejected. They find that employees who care most about growth and standing suppress risky deviant work in order to protect their image (moderated mediation index −0.0446). Shi and Zhang find that fear of losing credit drives people to hoard ideas. Bai and He show that the same structure can empower or exhaust people depending on what they expect from it. Siebers et al. use agent-based simulation, which models many individual decision-makers acting and reacting over time. They show that autonomy granted without cultural support turned retail staff into defensive rule-enforcers, and they argue that such effects appear only when behavior is simulated dynamically.
We borrow three things from this work. The first is the mechanism: penalty for deviation leads to defensive conformity. The second is a moderator: the more an agent's payoff depends on institutional approval, the stronger the suppression. The third is the method: simulate the agents' responses rather than solve for a static equilibrium. We do not claim that survey findings from Chinese firms transfer to machines. They tell us where to look, and the federated experiments decide.
On the federated side, three papers set up the problem. FedVal shows that standard defenses against malicious clients cannot tell "different good" from "different bad." Its examples are Multi-Krum, which keeps the updates closest to their neighbors, and coordinate-wise median, which takes the median of each parameter across clients. FedVal fixes this with a server-side validation score that upweights clients who improve weak classes. That fix needs a clean server dataset of at least 10 samples per class, which strict privacy settings may forbid. FedProx adds a penalty $\frac{\mu}{2}\|w-w^t\|^2$ that tethers each client's local model to the global one. Read through our lens, this is an explicit price on deviation. FACT designs penalties and a three-agent "sandwich" competition so that clients cannot free-ride by lying about their costs. Its rewards depend on the gains of other agents. FACT's proofs assume IID data and no collusion, and it models strategy only as cost reporting, never as how distinctive an update to send.
None of these papers lets a client choose how much of its unusual signal to reveal. That choice is the object of this proposal. DyPFL is closest in spirit, because it treats a client's willingness to personalize as a decision driven by utility. Its decision, however, is whether to personalize, not how much to conform when contributing to the shared model.
Core Claim
Let client $k$ compute its honest local update $\Delta_k$ from global model $w^t$. The client then submits $w^t + \alpha_k \Delta_k$, where $\alpha_k \in [0,1]$ is a shrinkage factor it chooses. Setting $\alpha_k = 1$ is full honesty. Setting $\alpha_k \to 0$ means the client looks like the consensus. The client picks $\alpha_k$ to maximize
$$u_k(\alpha_k) = A_k\big(w^{t+1}\big) + \beta_k\, R_k(\alpha_k),$$
where $A_k$ is the accuracy of the next global model on client $k$'s own data and $R_k$ is the reward the mechanism pays. The weight $\beta_k$ says how much the client cares about reward relative to model quality. It is our analog of the "status concern" moderator in Liu et al.
We predict three things. Under acceptance-based or distance-based rewards with outlier-rejecting aggregators, minority-data clients settle at $\alpha_k$ well below 1, and the shrinkage grows with $\beta_k$. Majority-data clients barely shrink, because their updates already sit near the consensus. Under rewards equal to the update's measured benefit to peers, minority clients settle near $\alpha_k = 1$.
Research Program
We will build an agent-based federated simulator in the spirit of Siebers et al. Each round, every client computes its update and then chooses $\alpha_k$ by best response. To do this it evaluates a grid of $\alpha$ values against the mechanism's known scoring rule, using a local simulation of the aggregation step. It then submits. We iterate until the population's $\alpha$ profile stabilizes, or we record the cycles if it never does.
We will cross aggregators with reward rules. The aggregators are FedAvg, FedProx at several values of $\mu$, Multi-Krum, coordinate-wise median, and FedVal. The reward rules are flat participation pay, acceptance-based pay (paid only if not rejected), distance-based pay, FACT-style decoupled peer-gain rewards, and our proposal.
Our proposal is a peer-validated contribution reward. Each round, a random subset of clients scores every candidate update by its effect on their own local loss. Each scorer returns only a clipped, noise-added scalar. The server pays each contributor the sum of positive peer gains, with extra weight on classes where the global model is weak, borrowing FedVal's bias-reducer idea. The server uses the same scores to weight aggregation, which lets it reject updates that hurt peers. No server validation set is needed. The cost is extra client-side evaluation and a new attack surface: scorers can lie or collude. We test this directly rather than assume it away.
Baselines
The naive reference is FedAvg with flat participation pay and $\alpha_k$ forced to 1. This is the honest-client world that every cited paper evaluates. It gives the ceiling on what non-strategic minority clients contribute when no attackers are present, and its collapse under attack gives the floor.
The strongest existing method is FedVal, run as both aggregator and reward signal. As reported, it raises FEMNIST lowercase-letter accuracy from 32% to 53% over FedAvg, and ACSIncome minority recall from 19% to 50%. It also converges with 80% malicious clients (FedVal). All of these numbers come from honest, non-strategic clients. We will run FedVal with its server validation set at the paper's minimum size, on the same partitions, model, round budget, and hyperparameter search as our method. We will run it once with honest clients and once with best-responding clients.
Beating FedVal does not mean beating it in the honest setting. It means matching its minority accuracy and attack robustness within a few points while using no server data, and keeping minority $\alpha_k$ at least as close to 1 under strategic play.
FACT is the strongest incentive baseline. It reports reducing model loss by over 4x on MNIST, over 3x on CIFAR-10, and 66% for resource-constrained hospitals on HAM10000, each relative to local training (FACT). We implement its decoupled peer-gain reward on the same tasks, under non-IID splits, which its proofs do not cover.
The ablation that isolates our key ingredient holds the aggregator fixed (Multi-Krum, then FedVal) and changes only the reward rule from acceptance-based to contribution-based. Any difference in equilibrium shrinkage is then caused by the incentive, not by the aggregator. A second ablation removes the weak-class weighting from the peer scores.
Why not just…?
Why not just use FedVal and be done? This is the most serious threat. FedVal's bias reducer already upweights clients who help weak classes, so a FedVal-based reward may give minority clients no reason to shrink. If our experiments show near-unit $\alpha_k$ under FedVal rewards, we will report it plainly. The contribution then narrows to two things: a diagnosis that the acceptance and distance rewards common in incentive design do induce conformity, and a peer-scored replacement for FedVal's server validation set. That replacement is valuable precisely where the file flags FedVal as inapplicable, namely strict privacy settings.
Why not just lower FedProx's $\mu$ or drop the robust aggregator? Loosening the leash removes the explicit price on deviation. It also removes protection: FedVal reports that standard defenses already fail beyond roughly 30–40% attackers, and undefended FedAvg is worse. Lowering $\mu$ also ignores the original reason for the penalty, client drift, under which FedProx gains 22% absolute accuracy over FedAvg with 90% stragglers. Our $\mu$ sweep combined with attacker fractions will show whether any single $\mu$ keeps both minority accuracy and robustness. We expect none does.
Why not just let minority clients personalize locally? Personalization keeps the rare knowledge at the client that holds it, so it never reaches other clients that see the rare classes occasionally. DyPFL itself shows that personalization helps only some clients. We measure global minority-class accuracy averaged over all clients holding any rare-class data, which is exactly where local-only knowledge fails. A personalized baseline (local fine-tuning on top of FedAvg) is included.
Why not just pay flat participation rewards? Flat pay removes the conformity incentive but invites free-riding, which FACT exists to stop. We include flat pay with cost-strategic clients to show the trade-off.
Proposed Experiments
We use FEMNIST from LEAF, partitioned naturally by writer, with lowercase letters as the minority, matching FedVal's setup. We also use PUMS ACSIncome, with minority-group recall as in FedVal, and CIFAR-10 under Dirichlet non-IID splits with label-flipping and scaling attackers at 0%, 20%, 40%, and 80%. HAM10000 tests a cross-silo, FACT-like hospital setting.
The metrics are minority clients' equilibrium shrinkage $\alpha_k$, minority-class accuracy or recall, overall accuracy, accuracy of the 10th-percentile client (following LEAF's percentile reporting), reward paid to attackers, and client-side evaluation cost per round.
Three results are decisive. First, if minority clients do not shrink appreciably under Multi-Krum with acceptance rewards, the conformity-tax claim is false and the project stops there. Second, shrinkage should increase monotonically with $\beta_k$ in that setting. This is the federated analog of the status-concern moderation, and a flat curve would undercut the organizational lens. Third, under our reward, a colluding block of scorers at 10% and 30% of clients will try to boost attackers. If collusion at 10% restores attacker influence above FedVal's level, the server-data-free claim fails and FedVal's validation set remains necessary.
References
- Influence Mechanism of Employee Playfulness Personality on Employee Creative Deviance
- Inclusive Leadership and Creative Territory Behavior: A Triple Interactive Moderating Effect Model
- The “double-edged sword” effect of shared leadership on employee voice behavior
- Understanding Retail Productivity by Simulating Management Practices
- FedVal: Different good or different bad in federated learning
- On the Convergence of Federated Optimization in Heterogeneous Networks.
- FACT or Fiction: Can Truthful Mechanisms Eliminate Federated Free Riding?
- Personalizing or Not: Dynamically Personalized Federated Learning with Incentives
- LEAF: A Benchmark for Federated Settings