Teams of robots that map a shared field, such as ocean temperature, terrain or a demand surface, increasingly do so without a central server. The server is a single point of failure: it can be jammed, can drift out of radio range, or can simply crash. A Gaussian process (GP), a Bayesian regression model that returns both a prediction and a calibrated uncertainty, is the usual model for such fields. Scalable versions compress each robot's data into a small fixed-size summary over a handful of "inducing points," which are representative locations that stand in for the full dataset. Robots can then exchange these summaries peer to peer. This compression creates a security problem that the field has not solved. A single faulty or malicious robot can send a summary claiming near-zero uncertainty, and every honest robot that sums it in will collapse onto a wrong answer. The one existing defense, trimming the most extreme predictions, needs a central cloud to sort everyone's outputs, which brings back the server the system was built to avoid.
The idea of this note is that a GP summary carries a certificate of its own plausibility, so a robot can check each neighbor's summary on its own instead of comparing it to a majority. Under a fixed kernel, the amount by which honest data can shrink posterior uncertainty is exactly determined, and it is bounded by the number of observations a robot could physically have collected. The posterior mean an honest robot reports also has a known distribution under the prior. Each robot therefore admits a neighbor's summary only if three things hold: it spends no more "information" than the sender's sensing rate allows, its mean is statistically plausible under the prior, and it predicts the receiving robot's own held-out observations well wherever the two overlap. Admitted summaries are deduplicated by origin, so gossip loops do not count the same data twice. The claim is that this gives bounded damage per attacker with no central sorter and no hard cap on the attacker fraction. It also avoids a failure mode of trimming: when robots cover different regions, honest robots legitimately disagree, and trimming discards them.
Core Claim
Two lines of prior work meet here without overlapping. Decentralized GP fusion removes the server but trusts every peer. GP-DDF+ fuses local summaries peer to peer and proves the result is equivalent to a centralized partially independent approximation. GP-DDF-ASS lets each robot keep its own support set and translates summaries between sets with bounded loss, scaling to 1,083,608 sea-surface-temperature points. RESIN prevents double counting over cyclic gossip graphs with a Chernoff-weighted product. None of the three defends against a peer that forges its mean or covariance, and the research file notes that forged means break the Chernoff weighting and collapse variance network-wide. Robust GP aggregation, by contrast, keeps the center. BtPoE trims the highest and lowest predictive means and variances in a cloud before a product-of-experts combination, tolerates under 25% Byzantine agents, and needs a global sort at every step. AsyncDGP handles stale experts but names a serverless design as future work.
This proposal is not a decentralized trimmed mean. It replaces "agree with the majority" with "be consistent with the GP prior, with physics, and with my data." The central quantity is the information a summary claims. Let $K_{uu}$ be the prior covariance of the $m$ inducing values, $\Sigma_j$ the posterior covariance claimed by sender $j$, $k_j$ the number of observations $j$ claims, $\sigma_f^2$ the kernel signal variance and $\sigma_n^2$ the noise variance. For an honest sender with a fixed kernel,
$$\tfrac12\log\det K_{uu}-\tfrac12\log\det\Sigma_j \;\le\; \tfrac{k_j}{2}\log\!\Big(1+\tfrac{\sigma_f^2}{\sigma_n^2}\Big),$$
The bound follows from three facts. A Gaussian posterior's entropy reduction does not depend on the observed values, so it equals the mutual information. Inducing values affect observations only through the function, so the information about them cannot exceed the information about the function. Hadamard's inequality then bounds $\tfrac12\log\det(I+K_{ff}/\sigma_n^2)$ by the sum of its diagonal terms. Because $k_j$ cannot exceed sensing rate times elapsed time, which is a physical quantity each robot can verify, a variance-collapse attack is ruled out by construction.
The mean $\mu_j$ gets a matching check. Under the prior, an honest posterior mean is distributed as $\mathcal N(0,\,K_{uu}-\Sigma_j)$, so
$$d_j=\mu_j^\top (K_{uu}-\Sigma_j)^{-1}\mu_j \sim \chi^2_m,$$
and a receiver rejects summaries whose $d_j$ falls far in the tail. A sequential held-out likelihood test then catches forgeries in regions where the receiver has its own data.
The hypothesis follows from these checks. An attacker that passes them can move an honest robot's posterior by at most one information budget. Damage therefore grows with the total adversarial budget relative to honest information, rather than failing abruptly at a fixed fraction of attackers.
Research Program
The system is built on additive summaries. As in GP-DDF, each robot's contribution to the posterior over the inducing values is a pair of natural parameters (precision and precision-weighted mean), and the fused posterior is the prior plus a sum of these pairs. Each robot keeps a bounded cache holding the latest signed, versioned summary from each origin. It gossips that cache asynchronously over whatever links currently exist, and it sums only admitted entries. Deduplicating by origin removes the rumor problem exactly rather than conservatively, and it needs no synchronized rounds or global clock. The design also has a cost: cache size grows with the number of origins in range. We will measure that cost, compare it with RESIN's constant-size messages, and test a k-hop cap.
We start with a shared inducing grid. We then move to agent-centric support sets, translated with the GP-DDF-ASS low-rank transfer, and check that the information bound survives translation. It should, since translation is a deterministic map that cannot add information. Kernel hyperparameters are fit once on a small data subset and then frozen. GPnn supports this choice, since it shows that cheap subset-based hyperparameter estimation loses little predictive accuracy at scale. We treat the frozen hyperparameters as a stated limitation rather than solving joint hyperparameter learning.
The analysis has two parts. The first bounds an honest agent's excess posterior error as a function of admitted adversarial information. The second derives the false-rejection rate of the three tests on honest summaries, which is set by the test thresholds.
The main remaining weakness is an adaptive attacker that reports plausible statistics with a small bias, placed in regions no honest robot observes. We will build exactly this attacker and report how much damage it does, rather than claim it is impossible.
Baselines
The naive reference points are plain peer-to-peer product-of-experts summation with no defense, and isolated local GPs that never communicate. The first shows the size of the threat. On synthetic same-value poisoning, BtPoE reports that undefended aggregation blows up to MSE above 26.5, against an attack-free MSE of about 0.0049 (BtPoE). The second sets a no-harm floor: a defense is useless if honest robots end up worse than they would be alone.
The strongest existing robust method is BtPoE run with its central cloud intact. On synthetic data it reports MSE of about 0.0236 under same-value poisoning. For Kin40k and SARCOS, the file gives no reported number on Lacuna for BtPoE. We give it an oracle that ignores its own single point of failure, so it is an upper reference rather than a competitor in the serverless setting.
The fairest serverless competitor is decentralized BtPoE. In this version, each robot applies the same β-trimming of means and variances to its own neighbors' predictions before a product-of-experts combination. We run it on identical graphs, data, attacks and kernel hyperparameters, and tune β on a validation split with the same budget used for our thresholds. RESIN's Chernoff fusion is the strongest benign decentralized fusion method. rBCM, which weights experts by information gain, is also included because it looks like a defense: it downweights weak experts, but an attacker can simply claim to be strong. For an attack-free accuracy ceiling we use centralized SVGP or partially independent approximations on all the data.
Three ablations each remove one ingredient: the information budget, the prior χ² test, and the held-out likelihood test. A fourth replaces origin deduplication with Chernoff weighting.
Our method beats the field if three things hold together. Its honest-agent standardized MSE and NLL must be no worse than decentralized trimming in the random-partition setting. It must be clearly better under spatial partitions. And it must never fall below the isolated-GP floor at attacker fractions where trimming collapses.
Why not just…?
Why not just run decentralized trimming (a local BtPoE) at every robot? This is the real threat, and it may work as well as our method when robots hold statistically similar data. Trimming flags an expert because it disagrees with its neighbors. In spatial deployments, honest robots disagree all the time: one far from a region reports the prior variance while another reports a confident estimate, and GP-DDF-ASS shows that local-only models degrade sharply at region boundaries. Trimming should therefore discard exactly the informative honest summaries. We settle this by sweeping the partition from random to spatially clustered with no attacker present. If trimming's false-rejection cost stays negligible under spatial partitions, the advantage of our checks is confined to variance-collapse attacks, and we will scope the claim to that.
Why not just use robust BCM's entropy weighting? rBCM weights an expert by how much it reduces uncertainty, so a forger that claims huge certainty receives the largest weight. That makes it the most vulnerable design under variance collapse. The ablation that keeps rBCM weights and drops the information budget will show this directly.
Why not just keep a central server with hot spares? Field multi-robot work, including planetary collaborative SLAM and multi-robot target search, operates where links to any fixed infrastructure break. ADBO shows a central manager dropping worker utilization below 25% at 1,920 workers. Spares do not help a robot that is out of radio range, so the node-kill experiment below compares the two directly.
Why not just skip communication and keep isolated GPs? Isolation is Byzantine-proof, but the cited fusion work (GP-DDF-ASS, RESIN, GP-DDF+) reports large accuracy losses from it. Isolation serves as our no-harm floor, not as a solution.
Proposed Experiments
The primary spatial benchmark is the Indian Ocean sea-surface-temperature dataset used by GP-DDF-ASS, with 64 agents. Each agent follows a trajectory that defines both its data and a time-varying communication graph with packet drops. The file reports constant communication overhead on this dataset but no RMSE, so we reproduce attack-free GP-DDF-ASS as the reference. Kin40k and SARCOS, the benchmarks BtPoE and AsyncDGP use, are partitioned both randomly and by k-means on inputs so that the effect of heterogeneity can be seen directly. The BtPoE synthetic setting is used to recover its reported numbers before any comparison. As a scale sanity check, we run attack-free airline delay data against rBCM's reported RMSE of 35.5 on about 5 million points.
The attacks are the following. Same-value and ±100 injections follow BtPoE. A variance-collapse attack claims near-zero covariance. Colluding attackers coordinate their forgeries. The adaptive small-bias attacker targets unobserved regions. Attacker fractions are swept from 0 to 50%. Metrics are honest-agent standardized MSE and NLL, the fraction of honest agents worse than isolation, the false-rejection rate on honest summaries, and bytes sent per agent.
Four checks decide the claim. First, with no attacker, our method's accuracy under spatial partitions must match undefended fusion while trimming loses accuracy. Second, honest-agent error must grow gradually with the attacker fraction past 25%, the point at which BtPoE's guarantee ends. Third, under variance collapse, RESIN and rBCM must fail while the information-budget ablation alone recovers most of the protection. Fourth, killing the highest-degree robot, or the oracle cloud, must leave our method's error essentially unchanged while centralized BtPoE stops producing fused predictions.
References
- Gaussian Process-Based Decentralized Data Fusion and Active Sensing for Mobility-on-Demand System
- Gaussian Process Decentralized Data Fusion Meets Transfer Learning in Large-Scale Distributed Cooperative Perception
- Rumor-robust Decentralized Gaussian Process Learning, Fusion, and Planning for Modeling Multiple Moving Targets
- Byzantine-tolerant federated Gaussian process regression for streaming data
- Asynchronous Distributed Gaussian Process Regression
- Leveraging Locality and Robustness to Achieve Massively Scalable Gaussian Process Regression
- Distributed Gaussian Processes
- Multirobot Decentralized Collaborative SLAM in Planetary Analogue Environments: Dataset, Challenges, and Lessons Learned
- Multi-Robot Target Search using Probabilistic Consensus on Discrete Markov Chains
- Asynchronous Decentralized Bayesian Optimization for Large Scale Hyperparameter Optimization