Cloud or edge? For securing AI, the question no longer holds. The future lies in adaptive, distributed security spanning the whole cloud – edge continuum. Research explains how to combine the power of the cloud with the speed of the edge, without compromising security.

When it comes to securing artificial intelligence, the debate has long been framed as a choice: should data and models be protected centrally, in the cloud, or locally, at the edge, where data is generated? That framing is now outdated. As AI moves into factories, hospitals, vehicles and energy grids, security increasingly needs both.

Two complementary lines of defence

“Cloud and edge computing are complementary,” says Djamel Khadraoui, Head of Reliable Distributed Systems Unit at the Luxembourg Institute of Science and Technology. “The cloud provides computing power, large-scale analysis and central coordination, while the edge enables fast local decisions, better privacy and continued operation when connectivity is limited. Together, they can make AI systems both more secure and more resilient.”

In practice, the two work as a continuum, each layer covering the blind spots of the other. An edge device can detect an anomaly and respond within milliseconds, isolating a compromised sensor or blocking a suspicious command. The cloud gathers these local observations and compares signals from thousands of devices to spot attack patterns that no single sensor could detect. It then improves its detection models and sends updated protections back to the entire fleet. If the connection to the cloud is lost, local intelligence keeps protecting critical operations, and the system resynchronises once connectivity returns.

A larger and more complex attack surface

“In distributed AI, the risk is also distributed,” Djamel Khadraoui explains. “Data, models and decisions move across many devices and locations, increasing the attack surface. The main challenge is to keep data protected, devices trustworthy, models secure and policies consistent, while preserving performance, privacy and operational continuity.”

Many edge devices operate in physically exposed locations, with limited computing power to run heavy security tools. They combine hardware, operating systems and network technologies from many vendors, making consistent policies hard to enforce. Keeping thousands of models up to date and verifying that they have not been tampered with requires robust orchestration. Performance is another constraint: security checks must not slow down applications that need to respond in real time, such as autonomous vehicles or industrial control systems.

AI also brings its own threats. Models can be manipulated through poisoned training data, fooled by adversarial inputs or probed to extract confidential information. When learning is spread across many nodes, a single compromised participant can corrupt the whole system. Organisations must also reconcile distribution with regulation, as the GDPR and the European AI Act demand transparency, accountability and control over where data is processed.

AI more secure, trustworthy and privacy-preserving

At LIST, we develop approaches that make AI more secure, trustworthy and privacy-preserving. Our work covers areas such as federated learning, secure distributed AI, cybersecurity, privacy protection and AI assurance, with the goal of enabling organisations to use AI without unnecessarily exposing sensitive data.
Djamel Khadraoui

Djamel Khadraoui, Head of Reliable Distributed Systems Unit, Luxembourg Institute of Science and Technology

Federated learning, for instance, allows models to be trained collaboratively across multiple sites without centralising raw data. This approach is particularly relevant for sensitive sectors such as health, finance and critical infrastructure. Two key research questions follow from it: making such distributed learning resilient against malicious participants, and ensuring that shared model updates reveal as little as possible about the underlying data. AI assurance, in turn, aims to provide evidence that AI systems behave as expected, a growing requirement under European regulation.

Another line of work concerns AI-driven threat detection across the cloud – edge continuum. It relies on lightweight models able to run on constrained devices and on self-adaptive mechanisms that adjust protection to context and risk. Zero-trust principles, in which no device or user is trusted by default, complete this toolbox.

These efforts are carried out with industrial partners and public actors in Luxembourg and across Europe, so that results can be tested in real-world conditions.

The future of AI security will be neither purely centralised nor fully decentralised. It will be adaptive, distributed and AI-enabled, spanning the entire continuum from cloud to edge. For organisations, the priority is to start designing their architectures accordingly today.