Zero-Trust and Reinforcement Learning for Secure Federated Intelligence in 6G Edge Networks

The introduction of $\mathbf{6 G}$ networks will bring unprecedented advancements in connectivity, intelligence and automation. However, this integration of AI also exposes 6 G networks to sophisticated cyber threats, including adversarial attacks, model poisoning and Sybil attacks. Traditional security mechanisms are inadequate against these threats, necessitating proactive security solutions. In response, this paper presents a zero-trust security framework tailored for decentralised AI inference in 6 G environments. It enforces continuous verification and identity-based access control by integrating Self-Sovereign Identity with ZeroKnowledge Proofs. Anonymity is preserved through zk-SNARKbased membership proofs, allowing edge nodes to authenticate without revealing their identities. Sybil resistance is achieved by registering cryptographic commitments, derived from unique node identifiers included in Verifiable Credentials issued by trusted authorities, as leaves in an on-chain Merkle tree. This ensures one-time registration of each legitimate node. To further strengthen trust and resilience, a reinforcement learning-based trust mechanism is deployed at the aggregator level to evaluate participating devices, facilitating the detection and isolation of malicious clients.