Abstract
Kosto is a decentralized framework designed to facilitate secure and fair outsourced computations. By integrating Intel SGX for hardware-level security with blockchain-based micro-payments, Kosto ensures data privacy and equitable compensation for compute nodes. The system employs instruction-level metering to prevent overcharging and a broker-mediated protocol to enable scalable, off-chain transactions. Experimental results demonstrate that Kosto maintains low performance overhead while providing robust guarantees for both clients and providers.
Introduction
Traditional cloud computing models rely on centralized trust, which introduces single points of failure and significant privacy risks for sensitive data. While decentralized alternatives like Golem have emerged, they often struggle with a critical security/efficiency trade-off: to verify results, they require redundant execution across multiple nodes, which is wasteful. Furthermore, these platforms frequently suffer from the "unfair exchange" problem, where a compute node may not be paid for partial work if a task is interrupted, or a client may refuse to pay after receiving a result. Kosto addresses these challenges by leveraging Trusted Execution Environments (TEEs) and a novel metering and payment protocol.
System Architecture
Kosto relies on three main pillars to ensure a functional marketplace: Trusted Execution Environments (TEEs), TEE-based metering, and a hybrid micro-payment protocol. By using Intel SGX, Kosto creates secure "enclaves" that protect data confidentiality and execution integrity, removing the need for redundant verification.
As shown in Figure 1, the architecture involves three primary entities: the Client (P), the Broker (B), and the Compute Node (C). The Broker acts as a matchmaker, solving a maximum bipartite matching problem to optimize network utilization. Crucially, Kosto implements delegated attestation, allowing the Broker to verify the hardware integrity of the compute node on the client's behalf. This enables the client to submit a task and go offline while the computation proceeds.
Fair Exchange and Metering Protocol
A core innovation of Kosto is its use of instruction counting as the "gold standard" for metering. Unlike wall-clock time or CPU cycles, which a malicious host could inflate to overcharge a client, instruction counts are deterministic and verifiable within the enclave. This metering logic is instrumented directly into the client's code.
To ensure fair exchange, Kosto decouples payment for computation from payment for result delivery. As illustrated in Figure 2, the protocol uses a series of cryptographic hashes and random values to facilitate micro-payments. The reward is split into a "computation portion," paid incrementally as work progresses, and a "delivery portion," paid only when the final result is handed over. This protects the compute node from losing all revenue due to a crash while incentivizing the final delivery of the output to the client.
Performance Evaluation
The authors evaluated Kosto using SPEC CPU2017 benchmarks and cryptographic throughput tests. Key findings include:
- Low Overhead: Computation-intensive tasks experienced overhead as low as 3% when running inside the SGX enclave. Instrumentation for instruction counting added a manageable 8% to 14% overhead.
- Scalability: The broker's matching algorithm proved highly efficient, matching 4,000 requests to 4,000 nodes in approximately 1.8 seconds.
- I/O Considerations: For I/O-intensive tasks, overhead reached up to 1.5x due to the costs associated with moving data across the enclave boundary.
Conclusion and Future Work
Kosto successfully demonstrates a decentralized marketplace that balances security, efficiency, and fairness. While the current prototype is limited to single-threaded applications and does not natively implement side-channel defenses, it provides a robust foundation for secure outsourced computing. Future work will focus on supporting multi-threaded workloads and exploring distributed matching algorithms to further enhance broker scalability.
Summary updated