This week we look at the engineering and security tradeoffs behind zkVM optimization, post-quantum signature aggregation, transcript specifications, and recursive privacy proofs. We also cover ZK tooling, proof provenance, integer SNARKs, verifiable FHE, VOLE, and private inference, with research on binary-field proximity gaps providing important context.
Anne-Laure Schmitt, Bruce Collie, Samuel Balco, Category Labs
Shows why native CPU optimizations can backfire inside a zkVM: doubling hash-table bucket fields to aligned 64-bit words reduces ZisK's proving cost, while selective Keccak memoization avoids caching one-off work. Also discusses adversarial collisions in cheap hash functions. Reports 57% lower COST than its initial implementation and 17% lower median proving time than ziskethone on 30 blocks with an RTX 5090; benchmark checks and manual review do not yet provide formal correctness guarantees.
https://www.category.xyz/blogs/optimizing-the-monad-client-for-zkvm-execution
Alberto Centelles, Quang Dao, Justin Thaler, Markos Georghiades, Michael Zhu, Andrew Tretyakov
Aerie pairs Falcon signatures with the Akita lattice PCS, offering an alternative to hash-based signature aggregation. Its M4 Max benchmark aggregates 32,768 Falcon-512 signatures in 645 ms with 16 threads; verification takes 53.6 ms on one thread. The important distinction is proof size versus transmitted data: per-signature salts bring the aggregate to 1,346 KiB, excluding public keys and messages. Common-salt mode reduces this to 66 KiB but requires a different signer and an additional unforgeability assumption. This is a research implementation; zero knowledge is not claimed.
https://www.a16zcrypto.com/posts/article/fast-post-quantum-signature-aggregation-with-lattice-snarks
Code and benchmark details: https://github.com/a16z/aerie
Matteo Frigo, abhi shelat
The new CFRG research-group draft specifies a Ligero-based ZK system with sumcheck. Its engineering value lies in the details implementations must agree on: typed, prefix-free transcript encoding, hash-and-expand state transitions, and challenge sampling without replacement. Useful for studying how protocol descriptions become interoperable implementations; this remains an Internet-Draft, not a finalized RFC.
https://datatracker.ietf.org/doc/draft-irtf-cfrg-longfellow/
ignotusnemo, Ethereum Research
Explores whether small and large recursive workloads can use different proof configurations instead of sharing one largest-size relation. An authenticated verifier-class bank carries a separate accumulator lane per class, trading flexibility for public state and terminal verification work that grow with the class count. The implementation follows one parent per step; merging multiple children and accounting for their dependencies remain open work for EIP-8288 aggregation.
ZK Podcast, Episode 414; Mathias Hall-Andersen and Giorgio Dell'Immagine
Discusses AI-assisted circuit optimization through zkGolf and Clean, a Lean framework for writing and verifying circuits. Covers the SHA-256 optimization experiment, competition exploits, and what formally checked circuit development could mean for ZK engineering and audits.
Quang Dao, Scott Duke Kominers, Justin Thaler, Kai Zhe Zheng
Recommend:
Gives Lean-verified counterexamples showing that beyond-Johnson proximity-gap results in large characteristic do not extend to binary fields. Covers domains used by Binius64, LeanVM, and Flock, with concrete limits on the soundness of a single proximity-gap check. Important when evaluating proposed proof-size reductions; these counterexamples constrain security analyses rather than establishing a break of every system using those domains.
Sebastian Angel, Sydnie-Shea Cohen, Eleftherios Ioannidis, Pratyush Mishra, Alex Ozdemir, Yuxiang Qiu, Alireza Shirzad
Recommend:
Compiles a unified protocol description into prover and verifier code, managing transcripts and Fiat-Shamir automatically; implements 35 proof systems. Its static analyses check completeness and special soundness at fixed parameter sizes, with current limitations including no automatic certification of zero knowledge or general knowledge soundness.
Wenhao Wang, Fan Zhang, Charalampos Papamanthou
Shows that inserting prover identity and a request nonce into Fiat-Shamir challenges can still permit precomputation attacks against proof provenance. Formalizes the requirement to perform fresh proving work and constructs provenance-binding SNARKs from code-based polynomial commitments, including BaseFold and FRI.
Jiseung Kim, Minkyu Kim, Changmin Lee, Hochang Lee
Refines hybrid algebraic attacks on regular syndrome decoding by coordinating guesses and variable elimination across blocks. Reports below-target security estimates for selected parameter sets used in Ferret, Wolverine, and committed VOLE, making this relevant to concrete parameter choices for code-based ZK and correlation-generation protocols.
Michael Adjedj, Geoffroy Couteau, Nikolaos Makriyannis
Builds maliciously secure non-interactive VOLE from a Paillier-based construction, reporting 535 bytes per entry and about 5 ms per party in single-threaded experiments. Applications include concurrently secure two-round two-party ECDSA; the paper also develops proofs of plaintext shortness for Paillier ciphertexts and integer commitments.
Marcel Keller, Sela Navot, Nirvan Tyagi
Shows how inconsistent global MAC keys across preprocessing batches, threads, or process runs can expose secrets through authentication checks. Demonstrates the issue in evaluated versions of MP-SPDZ, SCALE-MAMBA, and FANNG-MPC, and identifies state that implementations must preserve to enforce the protocols' key-consistency requirement.
Krzysztof Pietrzak, Leonid Reyzin
Proves communication lower bounds for efficient graph-labeling-based proofs of sequential work and gives a matching construction with maximum indegree two. Reduces the honest prover's total hash-input complexity from O(T log T) to O(T), while retaining asymptotically optimal communication and verifier hash-input complexity in this setting.
Kabir Peshawaria, Zeyu Liu, Ben Fisch, Eran Tromer
The revision adds a C++ implementation and replaces performance estimates with measured server, proof-size, and client costs. Combines blind interactive proofs with GKR to preserve BGV/BFV SIMD computation while adding integrity; one evaluated variant incurs 21–33× server overhead and 52–54 MB proofs.
Jessica Chen, Lucas Xia, Wilson Nguyen, Benedikt Bünz
Builds integer SNARKs through an integer polynomial commitment that supports evaluations modulo a prime chosen after commitment. The reported implementation proves an RSA accumulator benchmark 15–30× faster than circuit-based approaches and non-native Poseidon2 hashing 24× faster.
Liam Eagen, Ying Tong Lai
Encodes Groth16 proof invalidity so that a decoder can recover a designated secret when the proof is invalid. Combines this mechanism with projective garbling and BitVM for permissionless optimistic verification on Bitcoin, with about 21 kB of offline encoding in its BN254 instantiation.
Jinnuo Li, Shu Han, Yunyan Zou, Fan Zhang, Puwen Wei, Weijia Wang, Hao Cheng
Studies leakage of the secret witness through VOLE-correlation computations in FAEST and SDitH. Proposes six classes of differential power-analysis attacks, validated with simulated leakage and measurements on three ARM Cortex-M4 devices, highlighting implementation risks for signatures built from ZK techniques.
Brendan Funk, Jiafeng Xie
Presents an FPGA accelerator for FAEST, optimizing its dominant operations and their scheduling, including a dedicated L1-EM-fast design. Evaluates hardware cost and latency against existing signature accelerators; its fast variant reports at least 98% lower equivalent area-delay product than the compared complete SDitH design at the same security level.
Shweta Agrawal, Anuja Modi, Anshu Yadav, Shota Yamada
Gives a counterexample to the full evasive inner-product functional-encryption assumption of Hsieh, Lin, and Luo. Extends earlier attacks on pseudorandom functional encryption to the weaker inner-product setting, showing that the adversary's freedom to choose private coins remains a critical issue in these evasive-LWE assumptions.
Shweta Agrawal, Anuja Modi, Anshu Yadav, Shota Yamada
Constructs indistinguishability obfuscation from a new instance-independent, falsifiable, non-interactive lattice assumption involving LWE samples and a carefully structured short Gaussian matrix. The result relies on this additional assumption, with cryptanalysis informed by attacks on evasive LWE.
Hyeongmin Choe, Minsik Kang, Damien Stehlé
Examines FHE security when computation results are exposed, targeting parameter choices justified by heuristic correctness bounds. Gives bootstrapping-failure and noise/mask-skewing attacks, with experiments in HElib and OpenFHE; a BFV key-recovery attack takes minutes to hours for the tested parameters.
Shihe Ma, Jiaming Liu, Anyu Wang, Xiaoyun Wang
Proposes Generalized CKKS, trading SIMD slots for lower modulus consumption through a new encoding of high-precision values. Reports seven levels at 1,000-bit precision with ring dimension 2^14 and multiplications below 30 ms in its experiments, with applications to retaining precision after noise flooding.
Georg Fuchsbauer, Carter Luck, Adam O'Neill, Ian Rapko
Studies membership leakage in vector protections that preserve compatibility with existing search backends. Breaks Scale-and-Perturb without known plaintexts, analyzes how geometry leakage interacts with linked known plaintexts, and evaluates mitigations such as shuffling and secret projection.
If you're interested in our ZK Insights or have ideas for similar content, we highly encourage everyone to head over to our GitHub repo and submit a Pull Request. Join forces with like-minded ZKPunks to co-create!
✨ Web collection version: https://insights.zkpunk.pro/
✨ GitHub repo link: https://github.com/ZKPunk-Org/zk-insights

