ZKProof 5.5 in Barcelona was a blast! We focused on standardization,and all the 100 participants, well, participated! :slightly_smiling_face:

Here’s summary of the talks, for those who couldn’t make it, but also as reference for the workgroups we formed.

 

Daniel Kang gave a comprehensive overview (slides here) of the current capabilities of zero-knowledge proofs for machine learning (ZKML), clearly explaining what types of models like ImageNet, Twitter’s recommendation system, and GPT-2 can be proven today. He made a compelling case for the future potential of ZKML to enable trust in an increasingly digital world once efficiency improves, presenting benchmarks showing 50x faster proving with new hardware. By highlighting concrete applications like privacy-preserving inference and training audits and open sourcing frameworks, Kang is driving collaboration to advance ZKML and recruit participants to help make it practical at scale.

Overview:

  • Daniel gave a comprehensive technical overview of the capabilities of zero-knowledge proofs for machine learning (ZKML) today using frameworks like his open source ZKML.
  • He clearly explained the key applications that are practical now like privacy-preserving inference and outlined the advances needed to expand adoption.
  • By highlighting concrete benchmarks and use cases, he made a compelling case for the potential of ZKML once efficiency improves.

Current Capabilities:

  • Daniel discussed the guarantees of ZK proofs for ML, emphasizing how they enable trustless verification without interaction between parties.
  • He showed through benchmarks that today ZKML can handle models up to ImageNet scale for vision, production recommendation systems like Twitter’s, and small language models like GPT-2.
  • Daniel gave examples of applications like selective revealing of model weights/data for inference, training audits, generative model prompt marketplaces, and biometric ID.

Advances Needed:

  • He presented data on proving times today using CPUs and how new hardware like FPGAs can accelerate proofs 50x.
  • Daniel explained how innovations in proving systems like CQ and circuit optimizations will expand what’s possible.
  • He highlighted the need for collaboration to improve efficiency through frameworks like his open source ZKML.

Key Takeaways:

  • Daniel made a compelling case that ZKML will enable trust in an increasingly digital world.
  • But for mainstream adoption, advances are still needed to reduce proving costs, which he is driving through open source.
  • By showcasing concrete benchmarks and applications, he recruited participants to help make ZKML practical at scale.
  • This presentation built excitement for the potential of ZKML once efficiency catches up with the theory.