Accelerating Heterogeneous Agent Collaboration in Dynamic Edge Networks

This collection of recent networking and ML-systems papers presents advances with direct security, privacy, and safety relevance. Key contributions include PRADA, a two-stage offline-teacher distillation framework to accelerate heterogeneous edge–server LLM collaboration while avoiding online context uploads (lowering exfiltration risk but increasing reliance on locally distilled policies); GIRAF, a Governance-as-Code real-time risk-indexing framework for agentic 6G that formalizes verification staleness and exposes machine-readable telemetry and automated safety envelopes (useful for dynamic-