Ron Rothblum on Fiat-Shamir Attacks, Flock and SNARK.fast

Ron Rothblum discusses the implications of the Fiat-Shamir attack, emphasizing its role in converting interactive protocols into non-interactive ones. He warns that while this method can enhance efficiency, it may compromise the security of certain protocols, such as the GKR protocol. Rothblum has been researching this area for over a decade and believes that the success of theoretical work should not be judged solely by immediate practical applications.

His work with Sysinct focuses on applying theoretical concepts to practical zero-knowledge virtual machine (ZKVM) design, particularly through jagged polynomial commitments. He is currently developing Flock v2 and a binary field ZKVM, reporting that results from the Snark.Fast competition have exceeded his expectations. Rothblum highlights the critical role of polynomial commitment schemes in SNARK constructions, noting that many rely on error-correcting codes, with Reed Solomon being a common choice.

Rothblum suggests that chaining multiple hashes could mitigate the Fiat-Shamir attack, although he acknowledges that changes made by Polyhedra are only partial solutions. He calls for further research to either break the Fiat-Shamir method or enhance confidence in its security. He also notes that the diagonalization concept behind the attack was influenced by earlier research from 1994.

In his work on sketch code, Rothblum incorporates Reed Solomon for efficiency while ensuring message distinctness among similar inputs. He believes that combining different proof systems can improve both proving speed and verification efficiency, as indicated by promising benchmarks. Rothblum introduces Flock, a proof system designed for batch proving of Boolean computations, which has led to new ideas and follow-up work in the field.

He mentions that the Ethereum Foundation's decision to transition from Poseidon to a Boolean hash was driven by efficiency concerns, reflecting the ongoing evolution of cryptographic methods. Snark.fast is a collaborative effort aimed at enhancing Flock's efficiency, with a focus on accelerating the proving process while maintaining soundness to prevent false statements from being proven true.

Rothblum explains that the verifier in Flock ensures both completeness and soundness, which is crucial for developing faster provers without compromising security. However, he cautions that achieving completeness with Snark.fast is more challenging than ensuring soundness. He expresses concerns about the potential for AI to introduce malicious code into the prover, despite the performance improvements observed from crowdsourced AI projects.

AI has significantly optimized circuit performance, achieving two and a half to three times better results, particularly through effective GPU utilization. Rothblum notes that an Intel server without a GPU outperformed the original Flock implementation by over three times, showcasing AI's potential in experimenting with various computer components. The competition for proving systems involves testing new provers against the original Flock verifier, with ZK golf focusing on circuit optimization and soundness.

Looking ahead, Rothblum reveals that Flock v2 will incorporate engineering improvements and circuit changes to enhance proving performance, expanding support for more complex functionalities, including recursion. He anticipates that NVIDIA's addition of binary field support in their GPUs will further improve performance for DKVMs. Rothblum predicts that if proving can be made as fast as computing, it could revolutionize verifiable computation across all operations, including machine learning.

Currently, the overhead for proving hashes is about 200 times slower than computing them, an improvement from earlier expectations. Rothblum aims to reduce this overhead closer to 1x, which could unlock new applications that are currently unimaginable. He suggests that the need for ASICs for performance may be diminishing due to advancements in GPU capabilities for binary fields, indicating that the ability to verify computations could transform trust in software and computations performed by agents.