When phones and other edge devices train a shared model together (federated learning), the server only ever averages whichever devices happen to be online. Devices that are online more often then pull the model toward their own data, a distortion called objective inconsistency. The standard fix is to give rarely seen devices a larger say: multiply their update by roughly one over their participation rate. This proposal argues that this fix has a privacy cost nobody has measured. Federated systems usually hide each device's update with secure aggregation, a cryptographic protocol that reveals only the sum of updates, so each device hides "in the crowd." Inflating one device's weight by 100× makes that device dominate the sum. The crowd effectively shrinks to one, and an attacker who inverts the summed gradient back into images is in effect attacking that single rare device.
We propose to correct participation bias the other way around. Frequent devices randomly sit out, so that every device contributes at the same effective rate, and every update that enters the sum carries weight one. Upweighting is reserved for a bounded residual of very rare devices. Each device makes the sit-out decision locally from its own participation count. A device that sits out still submits a masked zero, so the server learns neither who abstained nor anyone's participation rate. The result is meant to be a debiased aggregator that works with unweighted secure aggregation and with standard differential-privacy noise, and that keeps the hiding crowd large in every round.
Core Claim
The privacy that secure aggregation gives a client in round $t$ is governed by how concentrated the aggregation weights are, not by how many clients nominally participate. A simple summary is the effective crowd size
$$N_{\text{eff}}(t) = \frac{\left(\sum_{i \in S_t} w_i^t\right)^2}{\sum_{i \in S_t} (w_i^t)^2},$$
where $S_t$ is the set of clients whose updates enter round $t$ and $w_i^t$ is the weight applied to client $i$'s update. $N_{\text{eff}}$ equals the cohort size when weights are equal and approaches 1 when one weight dominates. Inverse-probability debiasing, used by FedAU and by Debiasing FedAvg, drives $N_{\text{eff}}$ toward 1 exactly in the rounds where a rare client appears.
A nuance makes the claim sharper than the research file's framing. Inflating a lone client's update norm does not by itself help an attacker who already sees that update in plaintext, because the cosine-distance gradient matching used in inversion attacks is invariant to scale. The damage occurs when updates are summed. The rare client stops being hidden by the others.
Our method replaces importance weighting with importance thinning. Let $\hat p_i$ be client $i$'s own running estimate of how often it participates, and let $q$ be a target rate. When an online client is selected, it contributes its real update with probability $a_i = \min(1, q/\hat p_i)$, and it scales the update by $w_i = \max(1, q/\hat p_i)$, capped at $w_{\max}$. Every client with $\hat p_i \ge q$ then enters the sum at expected rate exactly $q$ with weight one. Only clients rarer than $q$ are upweighted, and by at most $q/\hat p_i$ rather than $1/\hat p_i$. The server picks $q$ for each round so that the predicted $N_{\text{eff}}$ stays above a floor $m$. A securely aggregated count of real contributors provides the averaging denominator. What is new is the combination of three properties: unbiasedness in expectation for all but the rarest clients, a per-round guarantee on crowd size, and compatibility with plain unweighted secure aggregation. The cost is that some updates from frequent clients are discarded, which increases sampling variance. Measuring that cost honestly is central to the program.
Relation to Prior Work
Four lines of work motivate the bias problem and supply the corrections we compete with. Debiasing FedAvg (Debiasing Federated Learning with Correlated Client Participation) models minimum-separation lockouts as a Markov chain and has clients rescale their step by $1/\hat p_i$. We borrow its client-side frequency counter but use it to decide whether to participate rather than how much to amplify. FedAU (A Lightweight Method for Tackling Unknown Participation Statistics in Federated Averaging) estimates participation on the server and caps weights with a cutoff interval. Its server-side weights are incompatible with unweighted secure aggregation, and its cap is a global constant rather than a per-round crowd constraint. MIFA (Fast Federated Learning in the Presence of Arbitrary Device Unavailability) and FedVARP (Fedvarp: Tackling the variance due to partial client participation in federated learning) instead cache every client's latest update on the server. That requires per-client plaintext updates, which is fundamentally at odds with secure aggregation, so we treat them as plaintext reference points rather than deployable competitors.
On the privacy side, generative-prior inversion (Gradient Inversion with Generative Image Prior) shows that a passive server can reconstruct images from gradients averaged over batches of up to 32. It supplies our attack: we run it on the secure-aggregation sum and ask whose data comes out. Beyond Inferring Class Representatives: User-Level Privacy Leakage From Federated Learning notes that intact secure aggregation blocks its user-level attack. Our thesis is that weight concentration silently reopens this door.
Two papers constrain the scope. LIA-SA (Breaking Secure Aggregation: Label Leakage from Aggregated Gradients in Federated Learning) shows that a malicious server tampering with batch-normalization layers can disaggregate the sum anyway. Our method does not defend against such active servers, and we claim protection only against honest-but-curious ones. OASIS (OASIS: Offsetting Active Reconstruction Attacks in Federated Learning) is complementary, a vision-only augmentation defense against active attacks.
Research Program
The first stage is measurement. On non-IID CIFAR-10 and SVHN, we simulate Bernoulli, Markovian, and cyclic participation as in FedAU's setup, plus the $R$-round lockout of Debiasing FedAvg. We run FedAvg, FedAU, and Debiasing FedAvg under unweighted secure aggregation; for FedAU we simulate oracle weighting inside the sum to isolate its privacy cost. We log $N_{\text{eff}}$ each round. We then apply generative-prior and pixel-space inversion to the summed update and score reconstructions against every participating client's data. The hypothesis is that, under inverse-probability debiasing, reconstructions in low-$N_{\text{eff}}$ rounds match the rare client's images. In contrast, FedAvg's sums yield blends.
The second stage builds the thinning aggregator and analyzes it. We will prove that the thinned update is unbiased for the population objective when every $p_i \ge q$, and we will bound the bias contributed by clients below $q$ under the cap $w_{\max}$. We will also characterize the extra sampling variance as a function of how spread out the $p_i$ are, extending the variance decomposition of FedVARP to thinned cohorts. Where the analysis allows, we will extend it to the Markov lockout model of Debiasing FedAvg using stationary rates. The third stage adds per-client clipping and Gaussian noise. Because every included update has weight at most $w_{\max}$, noise can be calibrated to a known sensitivity. Under $1/\hat p$ scaling, by contrast, clipping either erases the correction or forces noise scaled to the largest weight.
Baselines
The naive reference is FedAvg over whichever clients are online, run under unweighted secure aggregation. It has maximal $N_{\text{eff}}$ and uncorrected bias. Debiasing FedAvg reports 67.6% CIFAR-10 test accuracy for FedAvg at $R=5$ (Debiasing Federated Learning with Correlated Client Participation).
The strongest accuracy competitor under correlated participation is Debiasing FedAvg itself, which reports 73.8% on CIFAR-10 at $R=5$, against 68.0% for FedVARP in the same paper. It is also the strongest competitor on our specific question, because its client-side scaling is the only existing correction that already runs under unweighted secure aggregation.
FedAU is the strongest competitor under unknown Bernoulli, Markovian, and cyclic participation. No reported result was found on Lacuna for FedAU accuracy figures on these datasets, so we reproduce it with its published cutoff, including the cutoff $K=50$, and sweep the cutoff with the same budget we give our floor $m$. MIFA and FedVARP run in plaintext as an accuracy ceiling for memory-based correction, not as private competitors.
The key ablation is a weight-cap variant: inverse-probability weighting with a per-client cap chosen so that its realized $N_{\text{eff}}$ matches ours. This separates the benefit of thinning from the benefit of merely bounding weights.
For fairness, all methods share data partitions, participation traces, model architecture, local steps, number of rounds, and a fixed tuning grid of equal size for learning rate and for each method's single trade-off knob. Beating the field means matching Debiasing FedAvg's accuracy within one standard deviation across seeds, while raising the minimum per-round $N_{\text{eff}}$ and lowering rare-client reconstruction fidelity to FedAvg's level. Under a matched differential-privacy budget, we should exceed all debiasing baselines in accuracy.
Why not just…?
Why not just cap the weights, as FedAU's cutoff already does? This is the real threat. Capping bounds concentration, but it buys privacy by giving up debiasing for every client above the cap, with no per-round control over crowd size. Thinning instead removes bias for the frequent majority exactly, at the price of discarded work. Which trade-off is better is an empirical question, and the matched-$N_{\text{eff}}$ weight-cap ablation settles it directly. If the cap matches thinning on both accuracy and leakage, we will report that thinning adds nothing beyond its compatibility with secure aggregation and differential privacy.
Why not just add differential-privacy noise? The generative-prior attack paper reports that noise strong enough to stop such attacks severely hurts accuracy. More specifically for this setting, per-client clipping applied after $1/\hat p$ scaling undoes the correction, while clipping applied before it lets sensitivity grow with the largest weight. Our third-stage experiment compares accuracy at a fixed privacy budget across all aggregators to test whether bounded weights actually rescue utility.
Why not just rely on secure aggregation as is? Secure aggregation hides individual updates only if the sum is not dominated by one of them. Our first-stage measurement tests precisely whether debiased sums leak a single rare client. If reconstructions from FedAU-weighted and Debiasing-FedAvg sums turn out no more identifiable than FedAvg's, the motivation collapses, and we will say so.
Why not just enforce a large lockout $R$? Debiasing FedAvg shows that FedAvg becomes unbiased only at the maximal $R$, which amounts to a full round-robin over the population. That is impractical across millions of devices, and it does nothing for clients that are inherently rare.
Proposed Experiments
The decisive experiments use non-IID CIFAR-10 with the $R=5$ lockout, to compare against the 73.8% Debiasing FedAvg result, and SVHN, CIFAR-100, and CINIC-10 under FedAU-style participation down to 1% average availability. For each method we report final and worst-decile client accuracy and the distribution of per-round $N_{\text{eff}}$. We also report rare-client leakage: PSNR between generative-prior reconstructions of the summed update and the rarest participating client's images, together with the fraction of rounds in which that client is the best-matching source.
For calibration, the generative-prior inversion paper reports 15.58 dB versus 13.78 dB for pixel-space attacks on ImageNet at 64×64 for individual batches. A sum-level attack that approaches this range for rare clients would confirm the spotlight effect. We also sweep $q$ and $m$ to trace an accuracy-leakage frontier for thinning and for capped weighting.
The claim stands only if three things hold. Thinning's frontier must dominate the capped-weighting frontier. Its extra variance must not erase its accuracy gain within the fixed round budget. And it must preserve the accuracy gain of debiasing under a matched differential-privacy budget. Failure on the variance check, which is most likely at extreme heterogeneity in participation rates, would mark the regime where thinning should hand off to bounded upweighting.
References
- Debiasing Federated Learning with Correlated Client Participation
- A Lightweight Method for Tackling Unknown Participation Statistics in Federated Averaging
- Fast Federated Learning in the Presence of Arbitrary Device Unavailability
- Fedvarp: Tackling the variance due to partial client participation in federated learning
- Gradient Inversion with Generative Image Prior
- Beyond Inferring Class Representatives: User-Level Privacy Leakage From Federated Learning
- Breaking Secure Aggregation: Label Leakage from Aggregated Gradients in Federated Learning
- OASIS: Offsetting Active Reconstruction Attacks in Federated Learning