Artificial Intelligence (AI) is set to be the engine of the 6G revolution. From optimizing network traffic to managing cybersecurity, we will be reliant on AI-driven systems. But as more of these intelligent algorithms enter our lives, an important question arises:
Can we actually trust them?
This is the focus of Proof of Concept 7 (PoC#7) within the MARE project, titled “Trustworthy Operation of AI”. While other security measures protect the network from hackers, PoC#7 protects the network from unreliable intelligence, ensuring that the AI models making life-or-death decisions are accurate, ethical and uncompromised.
The Hidden Risk: When AI goes Rogue
In a complex 6G environment, AI models are not static, but continuously learn and evolve. However, they are also vulnerable. An AI model can “drift” over time, losing accuracy as real-world conditions change, or it can be poisoned by malicious data during training. Even worse, an AI component might technically “work” but make decisions that are biased or violate privacy rules.
If a network operator blindly trusts a compromised AI model, the consequences could be severe – from service outages to privacy breaches and beyond. The challenge is that traditional security tools scan for viruses, not for “bad decisions” or “model drift.”
How PoC#7 builds a Trust Score
PoC#7 introduces a Trust Scoring Engine that acts as a background check for every AI component in a network. It enforces a “Zero Trust” approach, meaning no AI is trusted by default. Instead, every model must earn its trust through continuous verification.
The solution evaluates AI based on three critical pillars:
- Technical Robustness: Is the AI performing accurately? The system runs continuous tests to ensure the model isn’t making confident guesses about data it doesn’t understand.
- Data Integrity: Is the data fueling the AI clean? Using various techniques, the system verifies that the datasets used for training haven’t been secretly tampered with or corrupted.
- Operational Reliability: Is the model behaving normally? A data sanitation pipeline detects outliers and anomalous entries that could indicate a glitch or an attack.
A "Credit Score" for AI
All these checks are aggregated into a single AI Trustworthiness Score – which acts as a credit score for software. If an AI model’s score drops below a certain threshold – perhaps because its accuracy is fading or its data source looks suspicious, the MARE Security Plane takes immediate action.
The Adaptive Security & Privacy Orchestrator (ASPO) can automatically quarantine the untrustworthy model, trigger a retraining process, or alert human operators. This “closes the loop,” ensuring that only high-quality, verified intelligence is allowed to steer the network.
Why This Matters for the Future
As 6G integrates into critical sectors like healthcare and autonomous transportation, we need more than just smart networks – we need accountable ones.
PoC#7 provides the framework to measure and manage that accountability. By validating the integrity of AI across its entire lifecycle – from data ingestion to final decision, MARE is building a foundation where we can confidently use the power of AI, knowing that a sophisticated digital safety net is always watching.
