Differentially-Private Federated Linear Bandits
Authors: Abhimanyu Dubey, Alex `Sandy' Pentland
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4.4 Experiments. Figure 1: A comparison of centralized FEDUCB on 3 different axes. Fig. (A) describes the variation in asymptotic per-agent regret for varying privacy budget ε (where δ = 0.1); (B) describes the effect of n in private (solid) vs. non-private (dashed) settings; (C) describes the effect of d in per-agent regret in the private setting (n = O(M log T), ε = 1, δ = 0.1). Experiments averaged over 100 runs. |
| Researcher Affiliation | Academia | Abhimanyu Dubey and Alex Pentland Media Lab and Institute for Data, Systems and Society Massachusetts Institute of Technology {dubeya, pentland}@mit.edu |
| Pseudocode | Yes | Algorithm 1 CENTRALIZED FEDUCB(D, M, T, ρmin, ρmax). Algorithm 2 PRIVATIZER(ε, δ, M, T) for any agent i |
| Open Source Code | No | No statement regarding open-source code availability or links to repositories was found. |
| Open Datasets | No | For any d, we randomly fix θ Bd(1). Each Di,t is generated as follows: we randomly sample K d2 actions x, such that for K 1 actions 0.5 x, θ 0.6 and for the optimal x , 0.7 x , θ 0.8 such that 0.1 always. yi,t is sampled from Ber( xi,t, θ ) such that E[yi,t] = xi,t, θ and |yi,t| 1. |
| Dataset Splits | No | The paper describes generating synthetic data for a sequential learning problem but does not specify traditional train/validation/test dataset splits. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, memory) used for running experiments were provided. |
| Software Dependencies | No | No specific software dependencies with version numbers were mentioned. |
| Experiment Setup | Yes | For all experiments, we assume L = S = 1. For any d, we randomly fix θ Bd(1). Each Di,t is generated as follows: we randomly sample K d2 actions x, such that for K 1 actions 0.5 x, θ 0.6 and for the optimal x , 0.7 x , θ 0.8 such that 0.1 always. yi,t is sampled from Ber( xi,t, θ ) such that E[yi,t] = xi,t, θ and |yi,t| 1. Results are in Fig. 1, and experiments are averaged on 100 trials. In this setting, we set n = O(M log T), d = 10 (to balance communication and performance), and plot the average per-agent regret after T = 10^7 trials for varying M and ε, while keeping δ = 0.1. |