Picture a group of phones, drones and sensors training one model together. They have no central server; each device passes its model updates to its radio neighbours. Two engineering choices are now routine in this setting. Weak devices train only a narrow slice of the network, and messages are routed around nodes whose batteries are low. Each choice is reasonable on its own. Both also change something nobody tracks: how many honest devices' updates have been summed together by the time a curious device first sees a given piece of the model. That count matters because an update seen alone can be inverted back into the training images that produced it. This is gradient inversion, an attack that searches for an input whose gradient matches the observed one. An update that has been averaged with many others is far harder to invert. Our central prediction is that the devices most exposed in such systems are not the weak ones. They are the strong ones whose wide "outer" model slice is shared with few or no peers. Energy-aware detours and physical movement can push them into that position.
We propose to measure leakage per parameter slice rather than per model, and to test whether this count predicts real reconstruction quality. If it does, we will use it as a constraint inside routing and participation decisions, so that longer battery life is not bought with silent privacy collapse. The work connects four threads. Gradient leakage is the core phenomenon. Architectural heterogeneity and energy-aware routing are the mechanisms that reshape it. Node mobility, which breaks the usual static-topology assumption, is the operating regime in which all three are evaluated.
Core Claim
Topology-Dependent Privacy Bound For Decentralized Federated Learning proves that the minimum information loss at a decentralized aggregation step scales roughly as $1/(2(m-1))$, where $m$ is the size of the honest connected component. When $m$ falls to one, reconstruction becomes near-perfect. This bound assumes a static graph and one shared architecture. We claim that under width-sliced training the correct quantity is a per-slice mixing count. For slice $s$ of client $i$'s update in round $t$, define $c_s(i,t)$ as the number of honest clients whose slice-$s$ updates are summed into the message carrying $i$'s contribution when that message first reaches a non-honest node.
In nested schemes such as HeteroFL and FjORD, every client updates the low-index "core" channels, but only high-capacity clients update the outer channels. Core slices are therefore well mixed and outer slices poorly mixed. A lone wide device in a neighbourhood effectively transmits its outer-slice gradient unmixed. Two further effects shrink outer-slice mixing for exactly these devices. Energy-aware routing in the style of An Energy-Efficient Multi-hop Routing Algorithm for Wireless Sensor Networks sends traffic through high-battery relays, which concentrates many unmixed messages at a few observers. Battery-driven participation throttling, as in REWAFL, shrinks the set of contributors in a round.
The research file flags the opposite hypothesis as open (Gap 1). It notes that ordered-dropout core neurons behave like leading singular directions, and Concealing Sensitive Samples against Gradient Leakage in Federated Learning shows that low-entanglement gradient subspaces aid inversion. Together these suggest narrow clients might leak more. We treat these as competing predictions and design the study so that the experiments decide between them.
Research Program
The first component is measurement. We build a decentralized gossip-learning simulator. In it, devices hold HeteroFL width levels ('a' down to 'e') or FjORD ordered-dropout widths. They move under a random-waypoint model or stay static. They harvest energy following the Elizabeth City solar trace and see the finite-state Rayleigh fading channels used in Decentralized Federated Learning With Energy Harvesting Devices. Messages travel multi-hop under one of three routing rules: shortest-path, residual-energy-weighted, or the relative-energy cluster-head weighting of A Weight Cluster-Based Hybrid Routing Algorithm of ZigBee Network.
Curious nodes are placed either at random or at the relays that energy-aware routing favours. They run three attacks:
- the cosine-similarity reconstruction of Inverting Gradients;
- its analytic inversion of fully connected layers;
- the peer-side functional-marginalization and echo attacks of On the (In)security of Peer-to-Peer Decentralized Machine Learning.
Because Beyond Inferring Class Representatives shows that observing a client over many rounds raises leakage, each adversary also accumulates observations across a window. Mobility enters as a two-sided variable. It may break the repeated observation of fixed neighbours that marginalization relies on, or it may expose a moving client to more distinct eavesdroppers. The file lists this as unresolved (Gap 3), and we measure it directly.
The second component is the analysis that tests the core claim. We log $c_s(i,t)$ for every slice and round, and ask whether its minimum predicts reconstruction quality better than the whole-model honest-component size $m$ used by Li et al.
The third component is a scheduler that treats mixing as a constraint. In each round it selects routes $r$ and participating slices to
$$\min_{r}\ \sum_{e \in r} E_e \quad \text{subject to}\quad c_s(i,t) \ge k \ \text{for every transmitted slice } s,$$
where $E_e$ is the transmission energy on link $e$ and $k$ is a target mixing count. When no route meets the target, the scheduler can do one of three things. It can delay a slice so that it merges with a later contributor. It can route the slice through a node that already holds a partial sum. Or, as a last resort, it can add noise to that slice only.
Coordination stays within two hops, the boundary that Decentralized Federated Learning With Energy Harvesting Devices found sufficient, because exchanging state more widely costs more energy than it saves.
Baselines
The naive reference point is decentralized gossip learning with HeteroFL-style slicing, shortest-path routing, and no privacy measure. It fixes the accuracy ceiling and the leakage floor.
The strongest energy competitor is residual-energy-weighted multi-hop routing combined with REWAFL-style battery-aware participation. It represents what a careful practitioner would deploy today. REWAFL reports reaching target accuracy 37.9% to 51.2% faster than Oort and saving between 18.2% and 54.9% of total energy. Those figures come from a star topology, so we will rerun the method in our mesh rather than borrow them.
The strongest privacy competitors are the following:
- Uniform local differential privacy. Calibrated noise is added to every transmitted update. Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks reports that at comparable protection this drops MNIST accuracy to roughly 78–85%. The same paper's quantization defense keeps 96.67% and pushes reconstruction PSNR (peak signal-to-noise ratio, where lower means a worse reconstruction) from 55.3 to 8.1 under the Deep Leakage from Gradients attack.
- The mixed-quantization defense itself. We include it so it can be tested against cosine-direction inversion, which it was not evaluated against.
- ERIS-style model sharding. ERIS shards updates across disjoint aggregators and reports that even two aggregators reduce reconstructions on LFW (a face-image dataset) to unrecognizable noise. In our setup the shards are carried over the same multi-hop energy-aware routes, so relays can see them in transit.
On the heterogeneity side, the reference numbers are from HeteroFL. Its CIFAR-10 non-IID mix of levels 'a'–'e' reaches 61.64% against 58.99% for FedAvg, from a centralized server. The attack reference is from Inverting Gradients, which reports about 14 dB PSNR against a trained ResNet-20.
Two ablations isolate the key ingredient:
- The k-threshold drop rule. Any slice with fewer than $k$ contributors is simply not sent. This uses the slice-level accounting with no routing change.
- A whole-model variant of our scheduler. It enforces Li et al.'s component size $m \ge k$ instead of per-slice counts.
All methods share the same data partitions, model family, simulated energy and channel traces, mobility seeds, adversary placement and attack budget, measured as optimizer iterations and restarts. Each method's own hyperparameters get the same tuning budget.
Beating the baselines means the following. Our scheduler must hold every client's worst-case reconstruction PSNR at or below the uniform-DP level. It must lose at most one to two points of accuracy relative to the naive reference. It must match the energy-aware baseline's network lifetime, measured as rounds until first node death or partition, within 10%. No baseline should meet all three conditions at once.
Why not just…?
Why not just add differential privacy noise to every update? Two papers in the file document its utility cost. The mixed-quantization paper reports the 78–85% versus 96.67% MNIST gap. The peer-to-peer analysis argues that local DP gives a worse privacy-utility trade-off than central DP. Uniform noise also spends its privacy budget on core slices that are already well mixed. Our ablation that adds noise only to under-mixed slices tests whether slice accounting recovers most of that lost utility. If uniform DP at matched PSNR loses only as much accuracy as our scheduler, the scheduler is not worth its complexity, and we will say so.
Why not just use secure aggregation or ERIS-style sharding? Pasquini et al. note that secure aggregation adds heavy overhead and can be evaded in peer-to-peer learning with frequent dropouts. Dropouts are exactly what battery throttling produces. ERIS is the more serious threat, since it removes leakage cheaply. However, it assumes non-colluding aggregators and direct logical links. Over a multi-hop energy-aware mesh, the high-battery relays that carry the shards become exactly the colluding observers it excludes. We run ERIS with shards routed over the same paths. If it still reduces reconstructions to noise at equal energy cost, then sharding plus our accounting is the better recommendation, and the scheduler becomes secondary.
Why not just drop poorly mixed slices (the k-threshold rule)? This is the simplest use of our own idea, and it may well be enough. That is why it is a primary ablation rather than a strawman. Its expected cost is that strong devices' outer slices rarely reach $k$ in sparse or mobile neighbourhoods, so wide capacity goes untrained and accuracy slides toward the small model. If the k-threshold rule matches the scheduler on accuracy and lifetime, the contribution reduces to the measurement result plus this rule. That is still a useful and publishable outcome.
Why not just train one small model everywhere? Doing so removes outer slices and their exposure entirely. HeteroFL, however, shows the accuracy cost of this "lowest common denominator" model on CIFAR-10 non-IID (58.99% versus 61.64%). It also shows that small models rise from 77% to over 90% when they take part in heterogeneous aggregation. FjORD reports gains of 1.53 to 34.87 points over extended federated dropout. We include the small-only model as a baseline so the privacy gained by giving up width is priced explicitly.
Proposed Experiments and Decisive Checks
We will use the following datasets:
- CIFAR-10 and CIFAR-100, with ResNet-style networks sliced HeteroFL-style and with FjORD ordered dropout;
- FMNIST in the 23-device, three-cluster partitioned configuration of UAV-Aided Decentralized Learning over Mesh Networks, where a moving relay rewires connectivity;
- LFW for face reconstruction.
The FMNIST setting has a published reference of about 75–80% accuracy for UAV relaying. On FMNIST, the two-hop energy-harvesting method reports 84.2% (IID) and about 80% (non-IID), which serve as the energy-constrained accuracy references.
Topologies are a static torus, a static random geometric graph, the same graph under random-waypoint mobility at three speeds, and the UAV-bridged clusters. Metrics are:
- reconstruction PSNR and SSIM (structural similarity) per client and per slice, reported as worst-case and mean;
- marginalization success for the peer adversary;
- test accuracy;
- energy per round;
- network lifetime.
The first decisive check is predictive. Across all conditions, the per-slice minimum $c_s$ must explain reconstruction PSNR clearly better than whole-model $m$, measured by rank correlation and by held-out regression error. If it does not, the core claim fails, and Li et al.'s bound is sufficient.
The second check is directional. It settles whether wide clients' outer slices or narrow clients' core slices leak more. Each answer leads to a different defense.
The third check concerns energy-aware routing as a cause. Moving the curious nodes from random positions to energy-favoured relays should measurably raise worst-case PSNR under residual-energy routing but not under shortest-path routing.
The fourth check is the Pareto comparison between our scheduler, the k-threshold rule, slice-only DP, uniform DP and multi-hop ERIS over accuracy, lifetime and worst-case PSNR, with mobility reported as a separate axis. That separation shows whether movement acts as a natural defense or as an extra exposure.
References
- Topology-Dependent Privacy Bound for Decentralized Federated Learning
- HeteroFL: Computation and Communication Efficient Federated Learning for Heterogeneous Clients
- FjORD: Fair and Accurate Federated Learning under heterogeneous targets with Ordered Dropout
- WDM: An Energy-Efficient Multi-hop Routing Algorithm for Wireless Sensor Networks
- REWAFL: Residual Energy and Wireless Aware Participant Selection for Efficient Federated Learning Over Mobile Devices
- Concealing Sensitive Samples against Gradient Leakage in Federated Learning
- Decentralized Federated Learning With Energy Harvesting Devices
- A Weight Cluster-Based Hybrid Routing Algorithm of ZigBee Network
- Inverting Gradients - How easy is it to break privacy in federated learning?
- On the (In)security of Peer-to-Peer Decentralized Machine Learning
- Beyond Inferring Class Representatives: User-Level Privacy Leakage From Federated Learning
- Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks
- ERIS: Enhancing Privacy and Communication Efficiency in Serverless Federated Learning
- UAV-Aided Decentralized Learning over Mesh Networks