PrivCirNet: Efficient Private Inference via Block Circulant Transformation
Authors: Tianshi Xu, Lemeng Wu, Runsheng Wang, Meng Li
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We compare Priv Cir Net with the stateof-the-art HE-based framework Bolt (IEEE S&P 2024) and HE-friendly pruning method Sp ENCNN (ICML 2023). For Res Net-18 and Vision Transformer (Vi T) on Tiny Image Net, Priv Cir Net reduces latency by 5.0 and 1.3 with iso-accuracy over Bolt, respectively, and improves accuracy by 4.1% and 12% over Sp ENCNN, respectively. |
| Researcher Affiliation | Collaboration | Tianshi Xu Peking University tianshixu@stu.pku.edu.cn Lemeng Wu Meta, Inc. lmwu@meta Runsheng Wang Peking University r.wang@pku.edu.cn Meng Li Peking University meng.li@pku.edu.cn |
| Pseudocode | Yes | Algorithm 1: Inverted Residual Fusion Algorithm |
| Open Source Code | Yes | Our code and checkpoints are available on Git Hub. |
| Open Datasets | Yes | We evaluate Priv Cir Net on Mobile Net V2 [28], Res Net-18 [55], and Vi T [29] across four datasets: CIFAR-10, CIFAR-100, Tiny Image Net and Image Net. |
| Dataset Splits | Yes | We evaluate Priv Cir Net on Mobile Net V2 [28], Res Net-18 [55], and Vi T [29] across four datasets: CIFAR-10, CIFAR-100, Tiny Image Net and Image Net. |
| Hardware Specification | Yes | All the experiments are performed on a machine with 2.4 GHz Intel Xeon CPU. For CIFAR and Tiny Image Net datasets, we train all models on a single NVIDIA RTX4090 GPU and a single NVIDIA A6000 GPU. For Image Net, we train all models on 8 NVIDIA A100 GPUs. |
| Software Dependencies | No | Priv Cir Net is built on top of the SEAL library [52] in C++. ... Following [7], we set n = 8192, security parameter λ = 128, plaintext bitwidth p = 41 and ciphertext bitwidth q = 218, which is also the default setting in SEAL library [52]. |
| Experiment Setup | Yes | Following [10, 53, 54], we simulate a LAN network setting via Linux Traffic Control, where the bandwidth is 384 MBps and the echo latency is 0.3ms. All the experiments are performed on a machine with 2.4 GHz Intel Xeon CPU. Following [7], we set n = 8192, security parameter λ = 128, plaintext bitwidth p = 41 and ciphertext bitwidth q = 218, which is also the default setting in SEAL library [52]. ... All baseline methods and Priv Cir Net are trained using identical hyper-parameters, including data augmentation, number of epochs, and others. These hyper-parameters are detailed in the configs folder within our codebase. |