The strongest conclusions from a doctoral project are often the boundaries around its results. Robust decentralized federated learning is promising, but it is also a moving target: topology changes, clients fail, modalities disappear and the information exchanged by peers can itself become sensitive.
The thesis connects three contributions, yet none of them should be read as a universal recipe. The survey maps an active field. Flighter evaluates a reliability mechanism in a controlled aerial-reconnaissance scenario. Modalis studies prototype-based multimodal collaboration on benchmark datasets. Together they identify a set of open questions that matter for real deployments.
1. From static graphs to living networks
Many evaluations use a fixed graph so that the learning mechanism can be isolated. Real systems do not stay fixed. A node can move out of range, a link can become congested or a new participant can join with a different clock and hardware profile.
The next step is to model availability and learning together. If is the mixing matrix at round , then a changing topology means that both the neighborhood and the effective consensus rate can vary:
where collects local learning updates. The research challenge is to distinguish a temporary graph problem from a poisoned or unreliable contribution without punishing useful participants for conditions outside their control.
2. Make context signals robust to manipulation
Situational-awareness indicators are useful only if their provenance and uncertainty are understood. Position, traffic, resource use and model similarity may be noisy, delayed or deliberately falsified. A future defense needs confidence intervals, missing-value policies and tests where the context channel is attacked at the same time as the model channel.
An operational score should therefore be accompanied by a reason and a validity window. “Low trust” is not enough. The protocol should be able to explain which signals changed, how much they affected the contribution and what recovery action is available.
3. Study privacy beyond data locality
Keeping examples on a device is not a formal privacy guarantee. Parameters, gradients and prototypes can leak information through inversion, membership or property-inference attacks. Prototype exchange deserves the same scrutiny as any other representation, especially when it is class-specific.
Promising directions include differential privacy, secure aggregation adapted to peer-to-peer settings and direct leakage evaluation. These mechanisms introduce their own costs, so quality, privacy budget and communication should be reported together rather than as isolated claims.
4. Test external validity
Controlled benchmarks are necessary for reproducibility, but they simplify sensor failures, mobility, task changes and operational constraints. Modalis uses four multimodal datasets and a specified client graph; Flighter uses a simulated mission with defined perturbations. The next evidence should include real or high-fidelity deployments, dynamic participation and cross-domain transfer.
External validity also means testing unfavorable cases: rare classes, clients with little support, stale prototypes, long disconnections and simultaneous faults. A method that works only when every assumption is satisfied is not yet a robust deployment technology.
5. Measure the whole system
Future studies should make the trade-off explicit:
| Property | Open question |
|---|---|
| Quality | Does the method preserve minority-client performance? |
| Resilience | How does it recover after a compromised peer? |
| Privacy | What can an observer infer from the exchanged object? |
| Efficiency | What is the cumulative communication and energy cost? |
| Governance | Who can change thresholds, weights or membership? |
| Reproducibility | Can another team recreate the topology and attack? |
This list is intentionally broader than accuracy. A DFL system is a socio-technical protocol: it combines models, networks, devices, policies and operators.
A measured research agenda
The most productive next experiments are not necessarily larger. They are experiments that relax one assumption at a time while keeping the comparison auditable: dynamic graphs, manipulated context, formal privacy mechanisms, real missing modalities and cross-dataset transfer. Each relaxation should state what it costs and what it improves.
That is the useful legacy of the thesis: decentralization, robustness and heterogeneity should be treated as connected design problems. The field will progress when results are reported with the same care as their limitations.
This note is an original synthesis of limitations and future directions identified by the doctoral thesis. It is not a reproduction of its conclusions.
