ExpProof : Operationalizing Explanations for Confidential Models with ZKPs
Authors: Chhavi Yadav, Evan Laufer, Dan Boneh, Kamalika Chaudhuri
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments. We evaluate Exp Proofon fully connected Re LU Neural Networks and Random Forests for three standard datasets on an Ubuntu Server with 64 CPUs of x86 64 architecture and 256 GB of memory without any explicit parallelization. Our results show that Exp Proofis computationally feasible, with a maximum proof generation time of 1.5 minutes, verification time of 0.12 seconds and proof size of 13KB for NNs and standard LIME. |
| Researcher Affiliation | Academia | 1UC San Diego 2Stanford University. Correspondence to: Chhavi Yadav <EMAIL>, Evan Monroe Laufer <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 LIME (Ribeiro et al., 2016), Algorithm 2 STANDARD LIME VARIANTS, Algorithm 3 BORDERLIME, Algorithm 4 FIND CLOSEST POINT WITH OPP LABEL, Algorithm 5 Exp Proof: Provable Explanation for Confidential Models, Algorithm 6 ZK LIME, Algorithm 7 ZK CHECK POSEIDON, Algorithm 8 ZK FIND OPP POINT, Algorithm 9 ZK LASSO, Algorithm 10 ZK TOP K, Algorithm 11 ZK EXPONENTIAL KERNEL, Algorithm 12 ZK UNIFORM SAMPLE, Algorithm 13 ZK GAUSSIAN SAMPLE. |
| Open Source Code | Yes | Our code is publicly available at : https:// github.com/emlaufer/Exp Proof. |
| Open Datasets | Yes | Datasets & Models. We use three standard fairness benchmarks for experimentation : Adult (Becker & Kohavi, 1996), Credit (Yeh, 2016) and German Credit (Hofmann, 1994). |
| Dataset Splits | No | The paper mentions evaluating results on '50 different input points sampled randomly from the test set' but does not provide explicit overall training, validation, and test split percentages or counts for the datasets used to train the models. |
| Hardware Specification | Yes | Our ZKP experiments are run on an Ubuntu server with 64 CPUs of x86 64 architecture and 256 GB of memory, without any explicit parallelization while ezkl automatically does multithreading on all the available cores, we do not use GPUs, do not modify ezkl to do more parallelization and do not do any of the steps in ZK LIME (Alg. 6) in parallel by ourselves. |
| Software Dependencies | Yes | We code Exp Proofwith different variants of LIME in the ezkl library (Konduit, 2024) (Version 18.1.1) which uses Halo2 (Zcash Foundation, 2023) as its underlying proof system in the Rust programming language, resulting in 3.7k lines of code. |
| Experiment Setup | Yes | Our neural networks are 2-layer fully connected Re LU activated networks with 16 hidden units in each layer, trained using Stochastic Gradient Descent in Py Torch (Paszke et al., 2019) with a learning rate of 0.001 for 400 epochs. ... Our random forests are trained using Scikit-Learn (Pedregosa et al., 2011) with 5-6 decision trees in each forest. ... We use the LIME library for experimentation and run the different variants of LIME with number of neighboring samples n = 300 and length of explanation K = 5. Based on the sampling type, we either sample randomly from a hypercube with half-edge length as 0.2 or a Gaussian distribution centered at the input point with a standard deviation of 0.2. Based on the kernel type we either do not use a kernel or use the exponential kernel with a bandwidth parameter as #features 0.75 (default value in the LIME library). Rest of the parameters also keep the default values of the LIME library. Our results are averaged over 50 different input points sampled randomly from the test set. The duality gap constant, ϵ is set to 0.001. |