New MARL Framework Cuts Privacy Leakage 85% for 6G VR Slices

Eighty-five percent less privacy leakage. That's the headline number from a new Multi-Agent Reinforcement Learning framework for VR slice management in 6G Software-Defined Radio Access Networks (SD-RAN). The authors claim a 34% throughput improvement and 28% reduction in resource consumption while keeping user data secure.

Why 6G VR Slices Need Privacy-Aware Agents

Virtual Reality in 6G demands end-to-end links with ultra-low latency and high throughput. SD-RANs must dynamically allocate network slices to meet those constraints, but existing methods either ignore mobility or leak user data when agents share information. This work tackles both problems together - agents collaborate without exposing raw user traces.

How Mobility Prediction and Information Bottleneck Work

Each agent predicts user movement to anticipate resource needs, then compresses shared state through an information bottleneck encoder. The encoder strips out privacy-sensitive features while retaining enough utility for coordinated decisions. The result: agents maximize resource distribution over VR links without exposing where a user is or what they are doing.

34% More Throughput, 28% Fewer Resources

Simulations compared against traditional baselines show the framework's MARL policy achieves 34% higher end-to-end throughput. Resource consumption drops by 28% because agents avoid over-provisioning. The 85% privacy leakage reduction comes from the bottleneck encoder filtering out identifiable patterns before they reach other agents.

None of that extra performance matters if the VR experience breaks. The authors report that latency and reliability stay within bounds for dependable immersive experiences, even under mobility. This framework sets a new target for secure, efficient VR slice management in 6G.

Source: Privacy-Aware Agent Collaboration for Dynamic VR Slice Management in 6G SD-RAN
Domain: arxiv.org