Can a Server Still Catch Free-Riders After Compression Hides the Gradients?

Federated learning trains one model across many devices: each device computes an update on its own data, sends it to a server, and receives the averaged model back. Two common engineering choices change what each upload looks like. Uploads are shrunk to save bandwidth, by zeroing all but the largest few percent of entries or rounding them to a few bits. The same shrinking is also recommended as a defense against gradient inversion, an attack that reconstructs a client's private training images from its uploaded update. A third concern pulls the other way. A rational client can "free-ride": skip the local computation, upload something that looks like an update, and still download the improved model. Every known free-rider detector inspects the uploaded vectors closely. Compression and privacy defenses exist to make those vectors less informative. Nobody has measured how badly they collide.

This note proposes to measure that collision and then to repair it cheaply. The core claim has two parts. First, a free-rider that replays the global model's last movement, compressed in the same way as honest uploads, becomes statistically indistinguishable from honest clients at roughly the compression levels that privacy work recommends. Second, detectability can be restored without the extra uplink traffic that current detectors need. The repair asks every client to include the same small, server-chosen random set of coordinates in its compressed upload, and scores clients by whether those coordinates agree with their peers on the part of the update that is new this round. The result would be a measured three-way frontier between bandwidth, reconstruction privacy, and free-rider detection, plus a detector that sits on that frontier rather than off it.

What prior work establishes and leaves open

The free-rider threat was formalized in Free-rider Attacks on Model Aggregation in Federated Learning. That paper shows that "disguised" free-riders, who add calibrated Gaussian noise to the received weights, look like ordinary statistical inliers, so poisoning-oriented anomaly detectors cannot see them. Its disguise is crude, though: pure noise carries no signal about the current descent direction. We study a stronger and equally cheap attacker that replays the most recent global update, $g_{t-1} = w_t - w_{t-1}$, where $w_t$ is the global model at round $t$, compresses it exactly as an honest client would, and adds noise. This attacker matters because momentum makes $g_{t-1}$ correlate strongly with every honest client's fresh update.

Two detectors set the reference point. Fed-PCA, from Data-Free Evaluation of User Contributions in Federated Learning, rewards correlation between clients' parameters on the same task and penalizes correlation across tasks. It needs no validation data, and it was tested against honest differential-privacy noise, but only on full-precision vectors and never against free-riders who hide behind that noise. WEF-Defense, from Rethinking the defense against free-rider attack from the perspective of model weight evolving frequency, tracks how often penultimate-layer weights cross dynamic thresholds. It requires clients to upload an extra matrix, and its authors name the resulting communication overhead as future work. Both papers motivate this proposal. Neither has been run on sparse or quantized uploads.

On the privacy and bandwidth side, two papers supply the compression regimes we must operate in and also constrain our design. Evaluating Gradient Inversion Attacks and Defenses in Federated Learning shows that pruning alone needs more than 99% sparsity to stop inversion. It also shows that 90% pruning combined with input encoding at batch size 32 reaches LPIPS 0.73, against 0.45 undefended (LPIPS is a perceptual distance where higher means a worse reconstruction), while keeping 86.10% CIFAR-10 accuracy. Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks keeps dequantization scales on the client and drops the PSNR of the DLG reconstruction attack from 55.3 to 8.1. Its server therefore cannot even reconstruct comparable real-valued updates, which removes the input that correlation-based detectors depend on. Any detector we propose must work under both regimes.

Reward-tiering work, Aequa, and truthful cost-reporting mechanisms, FACT or Fiction, decide what to pay each client once contributions are measured. They take a contribution signal as given. Our detector is meant to supply that signal when uploads are compressed.

Core mechanism

Each round, the server broadcasts a random seed. The seed determines an audit set $A_t$: a small set of coordinates, around 0.1% of parameters, that is the same for all clients. Each client sends its usual compressed payload. It also includes the values of its update on $A_t$, quantized with one public, shared scale, so the server can compare these values across clients even when private dequantization keys are in use. The server removes the predictable part of each audited update, meaning the projection onto the replayable global direction $g_{t-1}$, and keeps the remainder as the round's "innovation":

$$r_i = u_i|_{A_t} - \frac{\langle u_i|_{A_t}, g_{t-1}|_{A_t}\rangle}{\|g_{t-1}|_{A_t}\|^2}\, g_{t-1}|_{A_t},$$

Here $u_i$ is client $i$'s update and $|_{A_t}$ means restriction to the audit coordinates. Honest clients compute gradients at the same current model $w_t$, so their innovations share curvature-driven structure. A replay-plus-noise free-rider's innovation is essentially noise. Each client is scored by Fed-PCA-style peer agreement on $r_i$: the average correlation with other clients' innovations, minus the correlation expected between unmatched pairs drawn from different rounds. Unmatched pairs give an estimate of chance-level agreement that is free of the data distribution. This helps under non-IID data, where the file notes that existing detection margins narrow.

The design choices are each meant to be tested, not assumed. A shared audit set fixes the main failure of top-k uploads: different clients keep different coordinates, so their uploads barely overlap and cannot be compared. Removing the predictable part targets the strongest cheap attack. Randomizing the set each round stops a free-rider from computing gradients only on those coordinates. A per-coordinate gradient is not much cheaper than a full one, but whether partial computation is a profitable evasion strategy is an open question we will measure. Revealing $A_t$ only at upload time is the variant we will test against such an attacker.

Research Program

The work has three stages. The first is characterization. We cross compression level with free-rider strategy and run existing detectors unchanged on the decompressed uploads. Compression levels are top-k at 90%, 99%, and 99.9% sparsity, 8- and 4-bit quantization, and mixed quantization with client-held keys. Free-rider strategies are zero updates, Gaussian disguise, momentum replay, and replay plus noise matched to the honest variance. Detection quality is measured by area under the ROC curve (AUC) as the free-rider fraction varies from 15% to 90%, the range studied in the original attack paper.

The second stage builds the audit-and-innovation detector and runs ablations. We vary the audit-set size, remove the predictable-part step, and fix the audit set across rounds instead of redrawing it.

The third stage measures the privacy and bandwidth cost of the audit set itself. We run the strongest available inversion attack on the full upload, meaning the compressed payload plus the audit values. This shows whether a 0.1% dense slice gives an attacker meaningful help, measured as the change in LPIPS and PSNR relative to the compressed payload alone.

The output is a frontier plot with uplink bits per round on one axis, inversion LPIPS on another, and free-rider detection AUC on the third. Every method is placed on this plot.

Baselines

The naive reference point is FedAvg with no detection under each compression setting. It shows how much damage free-riders do when compressed. The prior attack paper suggests that this damage is mostly slower convergence, with large slowdowns only at around 90% free-riders. This baseline tells us whether the problem is worth solving at each operating point.

Fed-PCA on dequantized uploads is the strongest data-free competitor. It is the same family of peer-agreement statistic we use, so it directly isolates what the shared audit set and the removal of the predictable part add. Under client-held quantization keys, Fed-PCA will be given the client-side real-valued update as an upper-bound oracle, which is generous to the baseline. We tune its partitioning on the same validation rounds and budget as our method. No reported result found on Lacuna for Fed-PCA under sparsified or quantized uploads.

WEF-Defense is the strongest dedicated free-rider detector. It is run as published, with its extra matrix upload, and its additional bits are charged to its bandwidth. Its reported strength is a defense effect 1.68x and 1.33x stronger than CFFL and RFFL in IID and non-IID settings respectively, with 3.09x and 2.87x better stability up to 90% free-riders (Rethinking the defense…). Beating it means matching or exceeding its AUC at a strictly lower uplink cost.

Byzantine-robust aggregation, using coordinate-wise median and trimmed mean, is included as the reviewer's default. Cosine similarity to the aggregate serves as a one-line statistic.

Two ablations isolate our ingredients: innovation agreement on each client's own top-k support instead of the shared audit set, and shared-audit agreement on raw updates without removing the predictable part. Our method beats the field only if it reaches the highest detection AUC against the replay-plus-noise attacker at 99% sparsity, with uplink overhead under 1% and no statistically significant increase in inversion quality. All methods use identical data partitions, client counts, local epochs, and seeds.

Why not just…?

Why not just run Fed-PCA or WEF-Defense on the decompressed updates? This is the real threat. If Fed-PCA keeps high AUC at 99% sparsity, our detector is unnecessary and the contribution reduces to the characterization study. We expect it to fail for a concrete reason: different clients keep largely different top-k coordinates, so correlations between their sparse uploads are dominated by chance overlap. Under client-held quantization keys there is no comparable real-valued vector at all. The first stage tests exactly this, and we will report it either way. WEF-Defense may survive compression, but only by keeping its extra upload, which its own authors flag as a cost.

Why not just use Byzantine-robust aggregation? The original attack paper shows that disguised free-riders are inliers and therefore invisible to outlier filters (Free-rider Attacks…). A momentum-replay attacker sits even closer to the honest center. Including median and trimmed mean confirms this inside our own setup.

Why not just add differential privacy noise and skip compression-based privacy? The mixed-quantization paper reports that differential-privacy baselines fell to roughly 78–85% MNIST accuracy to match the protection that quantization gave at 96.67% (Mixed Quantization…). Privacy noise also gives free-riders a legitimate excuse for noisy uploads. We include a differential-privacy setting in the characterization so the frontier covers it.

Why not just pay by tier as Aequa does, or force honest computation with trusted hardware? Tiering still needs a contribution score, which is precisely what compression erodes. Trusted execution environments (Aequa) are absent on many edge devices and do nothing for bandwidth. Our detector is complementary: it can supply the score Aequa uses to assign widths.

Proposed Experiments

We use four benchmarks: CIFAR-10 with Dirichlet non-IID splits and a CNN, FEMNIST with natural device partitions, Shakespeare with an LSTM, and MNIST with LeNet-5 and 15 clients so we can match the mixed-quantization setting directly. Detection is reported as AUC and as final accuracy at fixed round budgets with 15%, 50%, and 90% free-riders. Privacy is measured by running the inversion attacks of Evaluating Gradient Inversion Attacks and Defenses and DLG at batch sizes 1, 8, and 32, reporting LPIPS and PSNR on the full upload including the audit set. Bandwidth is uplink bits per client per round.

Three checks are decisive. First, at 99% top-k sparsity against the replay-plus-noise attacker on non-IID CIFAR-10, existing detectors run on compressed uploads should drop toward chance AUC while ours stays well above it. If both stay high, the central motivation is false. Second, adding a 0.1% audit set should leave inversion LPIPS within noise of the payload-only value. If it does not, we report the privacy price of auditing as a negative result and shrink the audit set until it disappears or detection fails. Third, removing the predictable-part step should specifically collapse detection of the replay attacker while leaving detection of the Gaussian attacker intact. That pattern would show that the residual step, not the audit set alone, is what defeats the stronger attacker.