TrustFM 2027

Foundation models (FMs) – large pre-trained models for time series, vision, language and multi-modal data – are rapidly moving from data centers towards distributed smart systems. Yet five gaps keep FMs out of real deployments on the sensor-to-cloud continuum.

1. Physical consistency at the point of action. At the edge, FM outputs feed directly into actuators, grids and safety-relevant control loops, often without the validation and human oversight available in centralized settings. Purely data-driven models routinely violate conservation laws, boundary conditions and other domain constraints, and the quantization, pruning, early exiting and partitioning needed to fit models on edge devices can further degrade physical plausibility.
2. Data scarcity at deployment. Every new site, device or sensor modality brings its own distribution, while labeled data at the edge are scarce, costly or arrive only after a fault has occurred, so models must adapt from a handful of examples.
3. Trustworthiness, security and robustness. FMs are brittle under distribution shift, sensor faults, noise, adversarial perturbation, poisoning and privacy attacks, are opaque to operators, and seldom report calibrated uncertainty.
4. Efficiency. Billion-parameter models clash with the memory, energy and latency budgets of edge devices and constrained networks.
5. Distributed deployment. Heterogeneous cloud, edge and IoT resources require intelligent model partitioning, orchestration, routing and communication mechanisms to provide scalable and responsive inference.

TrustFM brings together the Artificial Intelligence, Distributed Systems, Edge Computing and IoT communities to address these gaps jointly, because in practice they are coupled. Compressing or early-exiting a model alters its calibration, robustness and physical plausibility; partitioning inference across edge nodes changes its failure modes; injecting physics priors can both enforce plausible outputs locally and reduce the data and compute a model needs, making it lighter to deploy; and meta-learned, low-shot adaptation decides how well all of these properties survive the move to a new site or device.

We therefore seek contributions that treat physical consistency, fast adaptation, robustness, explainability, uncertainty quantification and efficient, distributed inference as co-design objectives rather than post-hoc fixes.

Topics include, but are not limited to:

• Physics-informed and physics-guided foundation models
• Foundation models for sensor, time-series and multimodal data
• Model-agnostic meta-learning (MAML) for IoT
• Few-shot and low-shot adaptation
• Robustness and resilience of foundation models
• Explainability and interpretability
• Uncertainty quantification and calibration
• Adversarial robustness of foundation models
• Security of distributed and collaborative AI
• Efficient and adaptive inference
• Early-exit and conditional computation
• Model compression and efficient model optimization
• Distributed and collaborative inference
• Federated and privacy-preserving foundation models
• Resource-aware and hardware-aware FM deployment
• Adaptive caching and communication-efficient inference
• Energy-efficient and sustainable AI
• Digital sovereignty and trustworthy deployment
• Foundation models for cyber-physical and safety-critical systems
• Trust and risk assessment for AI systems

All deadlines are 23:59 Anywhere on Earth (AoE).

We invite original, unpublished contributions. Papers are selected primarily through an open call, complemented by a small number of invited papers from EU projects; all papers, invited or not, undergo the same peer-review process.

Regular papers

Up to 6 pages, IEEE conference format

Short / position papers

Up to 4 pages, IEEE conference format

Review

Single-blind, at least three reviews per paper by the Technical Programme Committee

Proceedings

Accepted papers will appear in the DCOSS-IoT 2027 workshop proceedings in IEEE Xplore

Templates

IEEE conference templates (LaTeX and Word)

TrustFM is planned as a half-day event, extendable to a full day depending on the number of accepted papers. The programme will open with an invited keynote by a recognised expert in the topics of the workshop.

Keynote speaker and detailed programme to be announced.

• Ioannis KontopoulosNational Technical University of Athens (NTUA-ICCS), Athens, Greece
• Antonios MakrisNational Technical University of Athens (NTUA-ICCS), Athens, Greece
• Konstantinos TserpesNational Technical University of Athens (NTUA-ICCS), Athens, Greece

The Technical Programme Committee will be announced soon.

For questions about the workshop, please contact the workshop chairs.

For registration, travel and venue information, see the DCOSS-IoT 2027 website.