Federated Unlearning for On-Device Recommendation

2025-05-06 0 0 2.77MB 9 页 10玖币

侵权投诉

Wei Yuan

The University of Queensland

Brisbane, Australia

w.yuan@uq.edu.au

Hongzhi Yin∗

The University of Queensland

Brisbane, Australia

h.yin1@uq.edu.au

Fangzhao Wu

Microsoft Research Asia

Beijing, China

wufangzhao@gmail.com

Shijie Zhang

Tencent

Shenzhen, China

julysjzhang@tencent.com

Tieke He

Nanjing University

Nanjing, China

hetieke@gmail.com

Hao Wang

Alibaba Cloud, Alibaba Group

Hangzhou, China

cashenry@126.com

ABSTRACT

The increasing data privacy concerns in recommendation systems

have made federated recommendations attract more and more at-

tention. Existing federated recommendation systems mainly focus

on how to eectively and securely learn personal interests and

preferences from their on-device interaction data. Still, none of

them considers how to eciently erase a user’s contribution to

the federated training process. We argue that such a dual setting

is necessary. First, from the privacy protection perspective, “the

right to be forgotten (RTBF)” requires that users have the right to

withdraw their data contributions. Without the reversible ability,

federated recommendation systems risk breaking data protection

regulations. On the other hand, enabling a federated recommender

to forget specic users can improve its robustness and resistance

to malicious clients’ attacks.

To support user unlearning in federated recommendation sys-

tems, we propose an ecient unlearning method FRU (Federated

Recommendation Unlearning), inspired by the log-based rollback

mechanism of transactions in database management systems. It

removes a user’s contribution by rolling back and calibrating the

historical parameter updates and then uses these updates to speed

up federated recommender reconstruction. However, storing all

historical parameter updates on resource-constrained personal de-

vices is challenging and even infeasible. In light of this challenge,

we propose a small-sized negative sampling method to reduce the

number of item embedding updates and an importance-based up-

date selection mechanism to store only important model updates.

To evaluate the eectiveness of FRU, we propose an attack method

to disturb federated recommenders via a group of compromised

users. Then, we use FRU to recover recommenders by eliminating

these users’ inuence. Finally, we conduct extensive experiments

on two real-world recommendation datasets (i.e. MovieLens-100k

∗Corresponding author.

Permission to make digital or hard copies of all or part of this work for personal or

classroom use is granted without fee provided that copies are not made or distributed

for prot or commercial advantage and that copies bear this notice and the full citation

on the rst page. Copyrights for components of this work owned by others than ACM

must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,

to post on servers or to redistribute to lists, requires prior specic permission and/or a

fee. Request permissions from permissions@acm.org.

Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

ACM ISBN 978-1-4503-XXXX-X/18/06. . . $15.00

https://doi.org/XXXXXXX.XXXXXXX

and Steam-200k) with two widely used federated recommenders to

show the eciency and eectiveness of our proposed approaches.

CCS CONCEPTS

•Information systems →Collaborative ltering.

KEYWORDS

Federated Recommender System, Machine Unlearning

ACM Reference Format:

Wei Yuan, Hongzhi Yin, Fangzhao Wu, Shijie Zhang, Tieke He, and Hao

Wang. 2023. Federated Unlearning for On-Device Recommendation. In

Proceedings of Make sure to enter the correct conference title from your rights

conrmation emai (Conference acronym ’XX). ACM, New York, NY, USA,

9 pages. https://doi.org/XXXXXXX.XXXXXXX

1 INTRODUCTION

Recommender Systems (RS) suggest the most appropriate items and

services to users by analyzing the collected personal data, e.g. user-

item interactions and user proles [

]. With the growing

awareness of privacy and the recent publishing of data privacy pro-

tection regulations such as the General Data Protection Regulation

(GDPR) [

] in the European Union and the California Consumer

Privacy Act (CCPA) [

] in the United States, collecting, storing,

and using users’ data is becoming harder. To address the above

challenges, more and more researchers focus on applying federated

learning [

] to recommendation systems (FedRecs), which train

recommender models on client devices without sharing user data

with a central server or other clients.

Since Ammad et al. [

] proposed the rst federated recommen-

dation framework, FedRecs have made great advancements re-

cently [

]. For example, Muhammad et al. [

] proposed FedFast

to accelerate training convergence. Lin et al. [

] investigated how

to exploit explicit feedback. Liang et al. [

] attempted to improve

the security of FedRecs via denoising techniques. Wu et al. [

]

incorporated GNN into a general FedRec framework.

Despite great advancements made in this area, it has been un-

explored how to forget specic users during the FedRec training

process. Without the ability to erase specic users’ contributions to

the federated training process, FedRec might break privacy protec-

tion laws or regulations such as CCPA and GDPR that give users

the right to control and withdraw their data at any time. Apart from

reducing the risks of breaking privacy protection rules, implement-

ing unlearning is also important to improve FedRec’s robustness

arXiv:2210.10958v2 [cs.IR] 3 Dec 2022

Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Wei Yuan, Hongzhi Yin, Fangzhao Wu, Shijie Zhang, Tieke He, and Hao Wang

and resistance to malicious attacks in an open setting where any

user/device can participate in the training process of FedRec. Re-

cent studies [

] have shown that current FedRecs are still not

“safe” enough when facing malicious users’ attacks. After detecting

such attacks, the ability to eciently erase such malicious users’

inuence without retraining from scratch is essential for FedRecs.

Although some recent works [

] tried to apply machine un-

learning to recommender systems because of data privacy concerns,

all of them focus on the traditional centralized recommenders. Their

methods require accessing the whole training data during unlearn-

ing, which is prohibitive in FedRecs. Some works [

–

] ex-

plored unlearning in federated learning, however, they are tailored

for classication tasks in Computer Vision (CV) area. To erase the

contributions of target clients, the most naive yet eective method

is to retrain the recommender model from scratch after removing

the target clients, which is infeasible in the real-world recommen-

dation setting due to its huge time and resource costs. Another

alternative is to continue the training after removing the target

clients. However, such a method cannot guarantee whether and

when these target users’ inuence on the global parameters (e.g.

item embeddings) will be erased. As a result, how to eectively

and eciently erase target clients’ contributions is not trivial in

FedRecs.

Inspired by the log-based rollback mechanism of transactions

in database management systems (DBMS), we propose to record

each client’s historical model updates. Once we need to erase some

users’ contributions, we will roll back and reconstruct the other

clients’ models according to their training logs (i.e. their histori-

cal model updates). To achieve that, the most intuitive way is to

keep all clients’ historical model updates at the central server. This

method can work when the number of clients is small, such as in

the federated classication setting in which there are only tens of

clients [

]. However, in FedRec, the number of clients is several

orders of magnitude larger than the classication settings. As the

storage costs at the central server increase linearly with the number

of clients, this naive method is unsustainable and impractical for

FedRec. Therefore, we propose to retain historical model updates

at each client’s local device, and storage costs at each client device

decouple with the number of clients. Still, it is non-trivial to store

each client’s historical model updates on the resource-constrained

device, and it is infeasible to simply store all updates.

In this paper, we propose FRU (Federated Recommendation

Unlearning), a simple yet eective federated recommendation un-

learning method. FRU is model-agnostic and can be applied to

most federated recommendation systems. The basic idea of FRU is

to erase a target client’s inuence by revising FedRec’s historical

updates and leveraging the revised updates to speed up FedRec

reconstruction. Compared with completely retraining (reconstruct-

ing) the FedRec from scratch, FRU requires less running time and

achieves even better model performance. FRU stores each client’s

historical updates locally on decentralized personal devices to avoid

high central server storage overhead. To eciently utilize the lim-

ited storage space on each client’s device, we design two novel

components: a user-item mixed semi-hard negative sampling com-

ponent and an importance-based update selection component. The

user-item mixed negative sampling exploits high-quality negative

samples to train FedRec, reaching comparable model performance

with fewer negative samples than the traditional sampling method.

Consequently, it reduces the size of item embedding updates at

each client. The importance-based update selection component dy-

namically chooses important updates to store on each client device

at each training epoch, instead of storing all parameter updates.

After achieving unlearning in FedRec, evaluating the eective-

ness of unlearning is neither easy because there are a large num-

ber of users in recommendation datasets, and many of them have

common items and similar preferences (actually, it is the base of

CF-based recommendation methods). Deleting a small portion of

normal users/clients will not signicantly change the performance

of FedRec. In light of this, we propose an attack method to destroy

the FedRec with a group of compromised clients/users (also called

malicious users). An eective unlearning method should recover the

destroyed FedRec quickly and achieve comparable or even better

performance than training without malicious users.

To demonstrate the eectiveness of our proposed approach, we

choose two commonly used recommenders [

], Neural Collab-

orative Filtering (NCF) [

] and LightGCN [

], with the most

basic federated learning protocol [

] as our base models. Then, we

conduct extensive experiments with these base models on two real-

world recommendation datasets, MovieLens-100k and Steam-200k.

The experimental results show that FRU can erase the inuence of

removed malicious users, with at least

x speedup compared with

the naive retraining from scratch.

The main contributions of this paper are summarized as follows:

•

To the best of our knowledge, this is the rst work to investi-

gate machine unlearning in federated recommender systems,

enabling FedRecs to eectively erase the inuence of specic

users/clients and eciently nish recovering.

•

We propose FRU, an unlearning method tailored for FedRecs.

It stores each client’s historical changes locally on their de-

vices. To improve the storage space eciency on resource-

constrained devices, we propose a novel negative sampling

method and an importance-based update selection mech-

anism. Then, FRU rolls back FedRecs to erase the target

users’/clients’ inuence and fast recover FedRecs by calibrat-

ing the historical model updates.

•

We design an attack method to intuitively evaluate the un-

learning eectiveness. The experimental results demonstrate

the eectiveness and eciency of FRU on two real-world

datasets. Comprehensive ablation studies reveal the eec-

tiveness and importance of each technical component in

FRU.

2 PRELIMINARIES

2.1 Federated Recommendation

Let

and

denote the sets of items and users (clients), respec-

tively. The sizes of items and users are

|V|

and

|U|

. Each user

𝑢𝑖

owns a local training dataset

D𝑖

, which contains user-item interac-

tions

(𝑢𝑖, 𝑣 𝑗, 𝑟𝑖 𝑗 )

𝑟𝑖 𝑗 =1

represents

𝑢𝑖

has interacted with item

𝑣𝑗

and

𝑟𝑖 𝑗 =0

means no interactions exist between

𝑢𝑖

and

𝑣𝑗

(i.e. neg-

ative samples). We denote the set of all negative instances for

𝑢𝑖

𝑛𝑒𝑔 (𝑖)

. The federated recommender is trained to predict the score

𝑟𝑖 𝑗

between

𝑢𝑖

and all non-interacted items. And then, according

to the predicted scores

𝑟𝑖 𝑗 ∈ [0,1]

, the federated recommender

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

FederatedUnlearningforOn-DeviceRecommendationWeiYuanTheUniversityofQueenslandBrisbane,Australiaw.yuan@uq.edu.auHongzhiYin∗TheUniversityofQueenslandBrisbane,Australiah.yin1@uq.edu.auFangzhaoWuMicrosoftResearchAsiaBeijing,Chinawufangzhao@gmail.comShijieZhangTencentShenzhen,Chinajulysjzhang@tencent.com...

展开>> 收起<<

Federated Unlearning for On-Device Recommendation.pdf

共9页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Federated Unlearning for On-Device Recommendation

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: