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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
DictPFL: Efficient and Private Federated Learning on Encrypted Gradients
Authors: Jiaqi Xue, Mayank Kumar, Yuzhang Shang, Shangqian Gao, Rui Ning, Mengxin Zheng, Xiaoqian Jiang, Qian Lou
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present Dict PFL, a practical framework that achieves full gradient protection with minimal overhead. It introduces two key modules: Decompose-for-Partial-Encrypt (De PE), which decomposes model weights into a static dictionary and an updatable lookup table — only the latter is encrypted and aggregated, while the static dictionary remains local and requires neither sharing nor encryption; and Prune-for-Minimum-Encrypt (Pr ME), which applies encryption-aware pruning to minimize encrypted parameters via consistent, history-guided masks. Experiments show that Dict PFL reduces communication cost by 402–748 and accelerates training by 28–65 compared to fully encrypted FL, while outperforming state-of-the-art selective encryption methods by 51–155 in overhead and 4–19 in speed. Remarkably, Dict PFL’s runtime is within 2% of plaintext FL, demonstrating for the first time that HE-based private federated learning is practical for real-world deployment. The code is publicly available at https://github.com/UCF-ML-Research/DictPFL. 5 Experimental Methodology, 6.1 Main Results, 6.2 Ablation Study |
| Researcher Affiliation | Academia | 1University of Central Florida 2Florida State University 3Old Dominion University 4University of Texas Health Science Center at Houston EMAIL EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the proposed methods, Decompose-for-Partial-Encrypt (De PE) and Prune-for-Minimum-Encrypt (Pr ME), using textual descriptions and mathematical formulas, but it does not include explicitly labeled pseudocode blocks or algorithms. |
| Open Source Code | Yes | The code is publicly available at https://github.com/UCF-ML-Research/Dict PFL. |
| Open Datasets | Yes | Datasets. We conduct experiments on three image classification tasks: CIFAR-10 [51], GTSRB [52], and Diabetic Retinopathy [53], as well as AG’s News [54] for sentence classification and Meta Math QA [55] for text generation. |
| Dataset Splits | Yes | We generate homogeneous data splits by randomly assigning training examples to individual clients without replacement. For heterogeneous settings, we simulate data heterogeneity by sampling the label ratios from a Dirichlet distribution with a symmetric parameter, following the [56]. In both settings, each client holds the same number of samples, following [57]. |
| Hardware Specification | Yes | All experiments were conducted on an AMD Ryzen Threadripper PRO 3955WX processor (2.2 GHz) with 125 GB of memory. |
| Software Dependencies | No | We adopt the CKKS homomorphic encryption scheme with bootstrapping [62–64], implemented via Open FHE [65]. The scheme is configured for 128-bit security following the Homomorphic Encryption Standard [66], with a cyclotomic ring dimension of N = 216, ciphertext modulus of 1555 bits, and multiplicative depth L = 12. Each ciphertext contains N/2 = 32,768 slots, enabling parallelized SIMD operations [67]. Data encoding follows the approach in [68]. While Open FHE is mentioned, no specific version number is provided for it or any other key software dependencies. |
| Experiment Setup | Yes | Hyperparameters. Unless otherwise specified, we set the dictionary size r to 4, the pruning ratio s% to 70%, the pruning patience τ to 3, and the reactivation probability scaler β to 0.2. Detailed analyses of these hyperparameters are provided in Section 6.2. |