Seng Pei Liew

Seng Pei Liew, Tsubasa Takahashi, Shun Takagi et al.

Recently, it is shown that shuffling can amplify the central differential privacy guarantees of data randomized with local differential privacy. Within this setup, a centralized, trusted shuffler is responsible for shuffling by keeping the identities of data anonymous, which subsequently leads to stronger privacy guarantees for systems. However, introducing a centralized entity to the originally local privacy model loses some appeals of not having any centralized entity as in local differential privacy. Moreover, implementing a shuffler in a reliable way is not trivial due to known security issues and/or requirements of advanced hardware or secure computation technology. Motivated by these practical considerations, we rethink the shuffle model to relax the assumption of requiring a centralized, trusted shuffler. We introduce network shuffling, a decentralized mechanism where users exchange data in a random-walk fashion on a network/graph, as an alternative of achieving privacy amplification via anonymity. We analyze the threat model under such a setting, and propose distributed protocols of network shuffling that is straightforward to implement in practice. Furthermore, we show that the privacy amplification rate is similar to other privacy amplification techniques such as uniform shuffling. To our best knowledge, among the recently studied intermediate trust models that leverage privacy amplification techniques, our work is the first that is not relying on any centralized entity to achieve privacy amplification.

Ryuichi Ito, Seng Pei Liew, Tsubasa Takahashi et al.

Applying Differentially Private Stochastic Gradient Descent (DPSGD) to training modern, large-scale neural networks such as transformer-based models is a challenging task, as the magnitude of noise added to the gradients at each iteration scales with model dimension, hindering the learning capability significantly. We propose a unified framework, $\textsf{LSG}$, that fully exploits the low-rank and sparse structure of neural networks to reduce the dimension of gradient updates, and hence alleviate the negative impacts of DPSGD. The gradient updates are first approximated with a pair of low-rank matrices. Then, a novel strategy is utilized to sparsify the gradients, resulting in low-dimensional, less noisy updates that are yet capable of retaining the performance of neural networks. Empirical evaluation on natural language processing and computer vision tasks shows that our method outperforms other state-of-the-art baselines.

Seng Pei Liew, Tsubasa Takahashi

We study Gaussian mechanism in the shuffle model of differential privacy (DP). Particularly, we characterize the mechanism's Rényi differential privacy (RDP), showing that it is of the form: $$ ε(λ) \leq \frac{1}{λ-1}\log\left(\frac{e^{-λ/2σ^2}}{n^λ} \sum_{\substack{k_1+\dotsc+k_n = λ; \\k_1,\dotsc,k_n\geq 0}}\binomλ{k_1,\dotsc,k_n}e^{\sum_{i=1}^nk_i^2/2σ^2}\right) $$ We further prove that the RDP is strictly upper-bounded by the Gaussian RDP without shuffling. The shuffle Gaussian RDP is advantageous in composing multiple DP mechanisms, where we demonstrate its improvement over the state-of-the-art approximate DP composition theorems in privacy guarantees of the shuffle model. Moreover, we extend our study to the subsampled shuffle mechanism and the recently proposed shuffled check-in mechanism, which are protocols geared towards distributed/federated learning. Finally, an empirical study of these mechanisms is given to demonstrate the efficacy of employing shuffle Gaussian mechanism under the distributed learning framework to guarantee rigorous user privacy.

Seng Pei Liew, Kenta Shinzato, Yuyang Dong

Modern Mixture-of-Experts (MoE) language models are designed based on total parameters (memory footprint) and active parameters (inference cost). However, we find these two factors alone are insufficient to describe an optimal architecture. Through a systematic study, we demonstrate that MoE performance is primarily determined by total parameters ($N_{total}$) and expert sparsity ($s:=n_{exp}/n_{topk}$). Moreover, $n_{exp}$ and $n_{topk}$ do not "cancel out" within the sparsity ratio; instead, a larger total number of experts slightly penalizes performance by forcing a reduction in core model dimensions (depth and width) to meet memory constraints. This motivates a simple principle for MoE design which maximizes $N_{total}$ while minimizing $s$ (maximizing $n_{topk}$) and $n_{exp}$ under the given constraints. Our findings provide a robust framework for resolving architectural ambiguity and guiding MoE design.

Seng Pei Liew, Satoshi Hasegawa, Tsubasa Takahashi

We study a protocol for distributed computation called shuffled check-in, which achieves strong privacy guarantees without requiring any further trust assumptions beyond a trusted shuffler. Unlike most existing work, shuffled check-in allows clients to make independent and random decisions to participate in the computation, removing the need for server-initiated subsampling. Leveraging differential privacy, we show that shuffled check-in achieves tight privacy guarantees through privacy amplification, with a novel analysis based on R{é}nyi differential privacy that improves privacy accounting over existing work. We also introduce a numerical approach to track the privacy of generic shuffling mechanisms, including Gaussian mechanism, which is the first evaluation of a generic mechanism under the distributed setting within the local/shuffle model in the literature. Empirical studies are also given to demonstrate the efficacy of the proposed approach.

Shun Takagi, Seng Pei Liew

Shuffling is a powerful way to amplify privacy of a local randomizer in private distributed data analysis, but existing analyses mostly treat the local differential privacy (DP) parameter $\varepsilon_0$ as the only knob and give generic upper bounds that can be loose and do not even characterize how shuffling amplifies privacy for basic mechanisms such as the Gaussian mechanism. We revisit the privacy blanket bound of Balle et al. (the blanket divergence) and develop an asymptotic analysis that applies to a broad class of local randomizers under mild regularity assumptions, without requiring pure local DP. Our key finding is that the leading term of the blanket divergence depends on the local mechanism only through a single scalar parameter $χ$, which we call the shuffle index. By applying this asymptotic analysis to both upper and lower bounds, we obtain a tight band for $δ_n$ in the shuffled mechanism's $(\varepsilon_n,δ_n)$-DP guarantee. Moreover, we derive a simple structural necessary and sufficient condition on the local randomizer under which the blanket-divergence-based upper and lower bounds coincide asymptotically. $k$-RR families with $k\ge3$ satisfy this condition, while for generalized Gaussian mechanisms the condition may not hold but the resulting band remains tight. Finally, we complement the asymptotic theory with an FFT-based algorithm for computing the blanket divergence at finite $n$, which offers rigorously controlled relative error and near-linear running time in $n$, providing a practical numerical analysis for shuffle DP.

Shokichi Takakura, Seng Pei Liew, Satoshi Hasegawa

In this paper, we study the problem of sampling from a distribution under the constraint of differential privacy (DP). Prior works measure the utility of DP sampling with density ratio-based measures such as KL divergence. However, such formulations suffer from two key limitations: 1) they fail to capture the geometric structure of the support, and 2) they are not applicable when the supports of the distributions differ. To deal with these issues, we develop a novel framework for DP sampling with Wasserstein distance as the utility measure. In this formulation, we propose Wasserstein Projection Mechanism (WPM), a minimax optimal mechanism based on Wasserstein projection. Furthermore, we develop efficient algorithms for computing the proposed mechanisms approximately and provide convergence guarantees.

Shun Takagi, Seng Pei Liew

We study $d$-dimensional unbiased mean estimation in the single-message shuffle model, where each user sends a single privatized message and the analyzer only observes the shuffled multiset of reports. While minimax-optimal mechanisms are well understood in the local differential privacy setting, the corresponding notion of optimality after shuffling has remained largely unexplored. To address this gap, we introduce the recently proposed shuffle index and use it to formulate the post-shuffling mechanism design problem as an explicit optimization problem. We then establish a minimax lower bound on the achievable mean squared error in terms of the shuffle index, which implies that mechanisms that are optimal under LDP can become suboptimal once shuffling is applied. Finally, we construct an asymptotically minimax optimal mechanism in the high privacy regime, which as a consequence achieves a privacy-utility trade-off nearly identical to that of the central Gaussian mechanism.

Seng Pei Liew, Takuya Kato, Sho Takase

Pretraining large language models (LLMs) is resource-intensive, often requiring months of training time even with high-end GPU clusters. There are two approaches of mitigating such computational demands: reusing smaller models to train larger ones (upcycling), and training computationally efficient models like mixture-of-experts (MoE). In this paper, we study the upcycling of LLMs to MoE models, of which the scaling behavior remains underexplored. Through extensive experiments, we identify empirical scaling laws that describe how performance depends on dataset size and model configuration. Particularly, we show that, while scaling these factors improves performance, there is a novel interaction term between the dense and upcycled training dataset that limits the efficiency of upcycling at large computational budgets. Based on these findings, we provide guidance to scale upcycling, and establish conditions under which upcycling outperforms from-scratch trainings within budget constraints.

Shokichi Takakura, Seng Pei Liew, Satoshi Hasegawa

The federated learning (FL) framework enables multiple clients to collaboratively train machine learning models without sharing their raw data, but it remains vulnerable to privacy attacks. One promising approach is to incorporate differential privacy (DP)-a formal notion of privacy-into the FL framework. DP-FedAvg is one of the most popular algorithms for DP-FL, but it is known to suffer from the slow convergence in the presence of heterogeneity among clients' data. Most of the existing methods to accelerate DP-FL require 1) additional hyperparameters or 2) additional computational cost for clients, which is not desirable since 1) hyperparameter tuning is computationally expensive and data-dependent choice of hyperparameters raises the risk of privacy leakage, and 2) clients are often resource-constrained. To address this issue, we propose DP-FedEXP, which adaptively selects the global step size based on the diversity of the local updates without requiring any additional hyperparameters or client computational cost. We show that DP-FedEXP provably accelerates the convergence of DP-FedAvg and it empirically outperforms existing methods tailored for DP-FL.

Seng Pei Liew, Takuya Kato

Bootstrapped pretraining, i.e., the reuse of a pretrained base model for further pretraining, such as continual pretraining or model growth, is promising at reducing the cost of training language models from scratch. However, its effectiveness remains unclear, especially when applied to overtrained base models. In this work, we empirically study the scaling behavior of bootstrapped pretraining and find that its scaling efficiency diminishes in a predictable manner: The scaling exponent with respect to second-stage pretraining tokens decreases logarithmically with the number of tokens used to pretrain the base model. The joint dependence on first- and second-stage tokens is accurately modeled by a simple scaling law. Such saturation effect reveals a fundamental trade-off in multi-stage pretraining strategies: the more extensively a model is pretrained, the less additional benefit bootstrapping provides. Our findings provide practical insights for efficient language model training and raise important considerations for the reuse of overtrained models.

Seng Pei Liew, Tsubasa Takahashi, Michihiko Ueno

We propose a new framework of synthesizing data using deep generative models in a differentially private manner. Within our framework, sensitive data are sanitized with rigorous privacy guarantees in a one-shot fashion, such that training deep generative models is possible without re-using the original data. Hence, no extra privacy costs or model constraints are incurred, in contrast to popular approaches such as Differentially Private Stochastic Gradient Descent (DP-SGD), which, among other issues, causes degradation in privacy guarantees as the training iteration increases. We demonstrate a realization of our framework by making use of the characteristic function and an adversarial re-weighting objective, which are of independent interest as well. Our proposal has theoretical guarantees of performance, and empirical evaluations on multiple datasets show that our approach outperforms other methods at reasonable levels of privacy.

Seng Pei Liew, Tsubasa Takahashi

Transfer learning is a useful machine learning framework that allows one to build task-specific models (student models) without significantly incurring training costs using a single powerful model (teacher model) pre-trained with a large amount of data. The teacher model may contain private data, or interact with private inputs. We investigate if one can leak or infer such private information without interacting with the teacher model directly. We describe such inference attacks in the context of face recognition, an application of transfer learning that is highly sensitive to personal privacy. Under black-box and realistic settings, we show that existing inference techniques are ineffective, as interacting with individual training instances through the student models does not reveal information about the teacher. We then propose novel strategies to infer from aggregate-level information. Consequently, membership inference attacks on the teacher model are shown to be possible, even when the adversary has access only to the student models. We further demonstrate that sensitive attributes can be inferred, even in the case where the adversary has limited auxiliary information. Finally, defensive strategies are discussed and evaluated. Our extensive study indicates that information leakage is a real privacy threat to the transfer learning framework widely used in real-life situations.