Cloud providers increasingly face the chal-lenge of selecting among rapidly evolving large language models (LLMs) for deployment at scale. Existing benchmarks, typically based on human or crowd preferences, advance model evaluation but rarely cap-ture the system-level constraints critical to cloud-based deployability. We present a system-aware LLM Arena benchmark that combines calibrated LLM-as-a-judge scores with runtime telemetry and cost signals. The benchmark measures judged utility together with end-to-end and tail latency, stability, memory usage, hard-ware utilization, and per-request cost and energy esti-mates. We further introduce a scene-aware composite score that makes latency-cost trade-offs explicit under real-time, budget-constrained, and balanced scenarios, reflecting the practical decision-making needs of cloud providers. Pareto frontiers visualize non-dominated models across these dimensions. Using locally deployed open-source models with DeepSeek-R1 as judge, the benchmark demonstrates how system factors reshape leaderboards compared to preference-only views, and provides actionable guidance for cloud/edge service providers to select and deploy LLMs under latency and cost constraints. For example, in our experiments GLM ranked highest in edge and government scenarios, but Qwen3_4B and Llama3_1B surpassed it in industrial and customer settings.

This paper proposes a new global model aggregation method based on using zero-knowledge federated learning (ZKFL). The purpose is to secure horizontal or P2P federated machine learning systems with shorter aggregation times, higher model accuracy, and lower system costs. We use a model parameter-sharing Chord overlay network among all client hosts. The overlay guarantees a trusted sharing of zero-knowledge proofs for aggregation integrity, even under malicious Byzantine attacks. We tested over popular datasets, Fashion-MNIST and CIFAR10, to prove the new system protection concept. Our benchmark experiments validate the claimed advantages of the ZKFL scheme in all objective functions. Our aggregation method can be applied to secure both rank-based and similarity-based aggregation schemes. For a large system with over 200 clients, our system takes only 3 seconds to yield high-precision global machine models under the ALIE attacks with the Fashion-MNIST dataset. We have achieved up to 85% model accuracy, compared to only 3%$\sim$45% accuracy observed with federated schemes without protection. Moreover, our method demands a low memory overhead for handling zero-knowledge proofs as the system scales greatly to a larger number of client nodes.