Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
On Probabilistic Truncation in Privacy-preserving Machine Learning
Authors: Lijing Zhou, Bingsheng Zhang, Ziyu Wang, Tianpei Lu, Qingrui Song, Su Zhang, Hongrui Cui, Yu Yu
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our benchmark indicates a 10x improvement in the DRe LU protocol, and a 6x improvement in the Re LU protocol over Piranha-Falcon and a 3.7x improvement over Bicoptor. As a result, the overall PPML model inference could be sped up by 3-4 times. [...] Our benchmark validates that in real-world performance, our technique consistently yields benefits, with only marginal accuracy losses incurred. In our experiments, we validate that employing our technique to scale 15-bit DRe LU does not affect the accuracy of our tested models. For applying 7-bit DRe LU in our technique, we assessed the impact on model accuracy as depicted in Fig 5 (see our full version). |
| Researcher Affiliation | Collaboration | 1Huawei Technology, China, 2The State Key Laboratory of Blockchain and Data Security, Zhejiang University, China, 3Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security, 4Shanghai Jiao Tong University, Shanghai |
| Pseudocode | Yes | Algorithm 1: The Truncation Protocol Proposed in Secure ML (Mohassel and Zhang 2017). [...] Algorithm 2: The Truncation Protocol Proposed in ABY3 (Mohassel and Rindal 2018). [...] Algorithm 3: The Non-interactive Deterministic Truncation Protocol. [...] Algorithm 4: The More General Non-interactive Deterministic Truncation Protocol. [...] Algorithm 5: Zero-preserving random mapping protocol. [...] Algorithm 6: Improved DRe LU Protocol. |
| Open Source Code | No | The paper cites 'Piranha source code. https://github.com/ucbrise/piranha/commit/dfbcb59d4e24ab69eb3606b49a102e602fdbee87. Accessed: 2014-10-25.' which refers to the code of a third-party platform (Piranha-Falcon) used for comparison, not the authors' own implementation or code release for the methodology described in this paper. There is no explicit statement from the authors about releasing their own code. |
| Open Datasets | Yes | Model Type Fraction precision ℓfrac x = 26 Secure ML Falcon Fantastic CIFAR10 Alex Net [...] CIFAR10 VGG16 [...] CIFAR10 Res Net18 |
| Dataset Splits | No | The paper mentions 'batch size 128' and refers to a 'calibration dataset that has the same distribution as the training dataset', but it does not provide specific training/test/validation dataset splits (e.g., percentages, sample counts, or explicit references to predefined splits). |
| Hardware Specification | Yes | We use three cloud server nodes to simulate three parties, each node with the following configuration: two Intel(R) Xeon(R) E5-2690 v4 @ 2.60GHz CPUs, 64 Gi B memory, and one independent Nvidia Tesla P100 GPU. |
| Software Dependencies | Yes | They are equipped with Ubuntu 16.04.7 and CUDA 11.4. |
| Experiment Setup | Yes | Finally, we finish our experiments in several DNN models with parameter ℓ= 64, ℓx = 7 or ℓx = 15. [...] We also simulate three different network environments: LAN1, LAN2, and WAN corresponded to 5Gbps/1Gbps/100Mbps bandwidth and 0.2ms/0.6ms/40ms round trip latency, respectively. [...] In particular, we monitor the range of the input values to the Re LU function. For each intermediate input data xi of the non-linear operator with the ith data in the calibration dataset, we denote the overall set of such input as X := {x0, . . . , x N 1}, where N is the size of calibration dataset. Let Max and Abs be the maximum and absolute functions, respectively. s is calculated by s = 2ℓx 1/(Max(Abs(X))) . |