Anthropic携手EPFL、MIT及StarkWare等机构,正式发布了名为FlashProve的系统,旨在通过零知识简洁论证(zkSNARK)技术解决大型语言模型的“可证明推理”问题 1。该系统的核心目标是确保用户收到的回复确实由声称的特定模型生成,从而防止路由服务替换模型或攻击者篡改Token流 1。FlashProve允许提供者在隐藏模型权重的前提下,证明响应是由承诺的模型生成的 1。
在性能方面,团队针对拥有4000亿参数的混合专家模型Llama 4 Maverick进行了测试,结果显示单张H100 GPU可实现每秒约110个token的证明生成速度 1。尽管这一过程比直接运行vLLM耗时高出70至95倍,但相比此前存在的开源系统,其速度提升了数个数量级 1。每处理一个1024-token的请求,生成的证明大小约为19 MB,而验证过程仅需普通CPU且耗时数秒 1。
FlashProve现已全面开源,包括博客文章、技术报告、协议规范以及基于C++和Lean验证器的可执行代码 1。为了保障安全,开发团队使用Lean证明助手对协议的完整性、可靠性和零知识性进行了形式化验证 1。相关文档均在Apache 2.0许可证下发布 1。该系统为AI发展节奏监控、入侵检测、文本溯源验证及合规模型审批提供了新的密码学工具 1。
Anthropic, in collaboration with EPFL, MIT, and StarkWare, has released FlashProve, a system designed to enable provable inference for large language models. 1 The core objective of this initiative is to address the inability of users to verify whether AI responses originate from specific claimed models, thereby preventing attacks where routing services swap models or adversaries tamper with token streams. 1 By leveraging zero-knowledge succinct arguments (zkSNARKs), FlashProve allows providers to demonstrate that a response was generated by a committed model without revealing the underlying model weights. 1
The system was tested on Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model. 1 On a single H100 GPU, FlashProve achieves a proof generation speed of approximately 110 tokens per second, which is several orders of magnitude faster than previous open-source systems. 1 For each request involving 1024 tokens, the resulting proof size is about 19 MB. 1 While the proof generation process takes roughly 70 to 95 times longer than running the model itself using vLLM, verification can be completed in seconds using standard CPUs. 1 To ensure robustness, the protocol’s integrity, soundness, and zero-knowledge properties have been formally verified using the Lean proof assistant. 1
The release includes a blog post, a technical report, protocol specifications, executable code in C++, and a Lean verifier, all documented under the Apache 2.0 license. 1 These tools provide new cryptographic mechanisms for monitoring AI development pacing, detecting intrusions, verifying text provenance, and facilitating compliance-based model approvals. 1
评论
还没有评论,欢迎留下第一条。