Yipeng Zhou

Miao Hu, Zhenxiao Luo, Amirmohammad Pasdar et al.

Edge computing has been getting a momentum with ever-increasing data at the edge of the network. In particular, huge amounts of video data and their real-time processing requirements have been increasingly hindering the traditional cloud computing approach due to high bandwidth consumption and high latency. Edge computing in essence aims to overcome this hindrance by processing most video data making use of edge servers, such as small-scale on-premises server clusters, server-grade computing resources at mobile base stations and even mobile devices like smartphones and tablets; hence, the term edge-based video analytics. However, the actual realization of such analytics requires more than the simple, collective use of edge servers. In this paper, we survey state-of-the-art works on edge-based video analytics with respect to applications, architectures, techniques, resource management, security and privacy. We provide a comprehensive and detailed review on what works, what doesn't work and why. These findings give insights and suggestions for next generation edge-based video analytics. We also identify open issues and research directions.

Laizhong Cui, Xiaoxin Su, Yipeng Zhou

Recently, blockchain-based federated learning (BFL) has attracted intensive research attention due to that the training process is auditable and the architecture is serverless avoiding the single point failure of the parameter server in vanilla federated learning (VFL). Nevertheless, BFL tremendously escalates the communication traffic volume because all local model updates (i.e., changes of model parameters) obtained by BFL clients will be transmitted to all miners for verification and to all clients for aggregation. In contrast, the parameter server and clients in VFL only retain aggregated model updates. Consequently, the huge communication traffic in BFL will inevitably impair the training efficiency and hinder the deployment of BFL in reality. To improve the practicality of BFL, we are among the first to propose a fast blockchain-based communication-efficient federated learning framework by compressing communications in BFL, called BCFL. Meanwhile, we derive the convergence rate of BCFL with non-convex loss. To maximize the final model accuracy, we further formulate the problem to minimize the training loss of the convergence rate subject to a limited training time with respect to the compression rate and the block generation rate, which is a bi-convex optimization problem and can be efficiently solved. To the end, to demonstrate the efficiency of BCFL, we carry out extensive experiments with standard CIFAR-10 and FEMNIST datasets. Our experimental results not only verify the correctness of our analysis, but also manifest that BCFL can remarkably reduce the communication traffic by 95-98% or shorten the training time by 90-95% compared with BFL.

Dongyuan Su, Yipeng Zhou, Laizhong Cui

Recently, federated learning (FL) has received intensive research because of its ability in preserving data privacy for scattered clients to collaboratively train machine learning models. Commonly, a parameter server (PS) is deployed for aggregating model parameters contributed by different clients. Decentralized federated learning (DFL) is upgraded from FL which allows clients to aggregate model parameters with their neighbours directly. DFL is particularly feasible for vehicular networks as vehicles communicate with each other in a vehicle-to-vehicle (V2V) manner. However, due to the restrictions of vehicle routes and communication distances, it is hard for individual vehicles to sufficiently exchange models with others. Data sources contributing to models on individual vehicles may not diversified enough resulting in poor model accuracy. To address this problem, we propose the DFL-DDS (DFL with diversified Data Sources) algorithm to diversify data sources in DFL. Specifically, each vehicle maintains a state vector to record the contribution weight of each data source to its model. The Kullback-Leibler (KL) divergence is adopted to measure the diversity of a state vector. To boost the convergence of DFL, a vehicle tunes the aggregation weight of each data source by minimizing the KL divergence of its state vector, and its effectiveness in diversifying data sources can be theoretically proved. Finally, the superiority of DFL-DDS is evaluated by extensive experiments (with MNIST and CIFAR-10 datasets) which demonstrate that DFL-DDS can accelerate the convergence of DFL and improve the model accuracy significantly compared with state-of-the-art baselines.

Jiating Ma, Yipeng Zhou, Qi Li et al.

To preserve the data privacy, the federated learning (FL) paradigm emerges in which clients only expose model gradients rather than original data for conducting model training. To enhance the protection of model gradients in FL, differentially private federated learning (DPFL) is proposed which incorporates differentially private (DP) noises to obfuscate gradients before they are exposed. Yet, an essential but largely overlooked problem in DPFL is the heterogeneity of clients' privacy requirement, which can vary significantly between clients and extremely complicates the client selection problem in DPFL. In other words, both the data quality and the influence of DP noises should be taken into account when selecting clients. To address this problem, we conduct convergence analysis of DPFL under heterogeneous privacy, a generic client selection strategy, popular DP mechanisms and convex loss. Based on convergence analysis, we formulate the client selection problem to minimize the value of loss function in DPFL with heterogeneous privacy, which is a convex optimization problem and can be solved efficiently. Accordingly, we propose the DPFL-BCS (biased client selection) algorithm. The extensive experiment results with real datasets under both convex and non-convex loss functions indicate that DPFL-BCS can remarkably improve model utility compared with the SOTA baselines.

Zhe Yuan, Yipeng Zhou, Jinghan Li et al.

Reinforcement learning with verifiable rewards (RLVR) has become an effective paradigm for improving reasoning language models on tasks such as mathematics, coding, and scientific question answering. However, widely used group-relative objectives, such as GRPO, summarize each sampled group with scalar statistics and therefore discard fine-grained relational information among candidate responses. This weakens credit assignment under sparse outcome rewards, especially when multiple generated solutions differ only subtly in reasoning quality. We propose \textbf{LamPO}, a \textbf{Lambda-Style Policy Optimization} method that replaces scalar group advantages with a \emph{Pairwise Decomposed Advantage}. LamPO aggregates pairwise reward gaps within each response group and modulates each comparison by a confidence-aware weight computed from sequence log-probability differences, while retaining the critic-free and clipped-update structure of PPO-style optimization. When reference solutions are available, we further add a lightweight ROUGE-L-based dense auxiliary reward to reduce reward sparsity. Experiments on AIME24, AIME25, MATH-500, and GPQA-Diamond with Qwen3-1.7B, Qwen3-4B, and Phi-4-mini show that LamPO consistently improves over GRPO and recent RLVR variants, with more stable training dynamics and better sample efficiency.

Zhe Yuan, Yipeng Zhou, Jinghan Li et al.

Group Relative Policy Optimization(GRPO) has become a cornerstone of modern reinforcement learning alignment, prized for its efficacy in foregoing an explicit value-critic by leveraging reward normalization across sampled trajectory cohorts. However, the method's reliance on a monolithic statistical baseline, such as the group mean, collapses the relational topology of the trajectory space into a single scalar, thereby erasing the fine-grained preference information essential for navigating complex, rank-sensitive reward landscapes. To address this issue, we introduce a novel framework, Lambda Policy Optimization (LambdaPO), that addresses this information-theoretic bottleneck by re-conceptualizing advantage estimation from a scalar value to a decomposed, pairwise preference structure. Specifically, the advantage for any given trajectory is formulated as the integrated sum of reward differentials against all peers in its cohort, where each pairwise comparison is dynamically attenuated by the policy's own probabilistic confidence in the established preference. To further mitigate the sparsity of binary outcome supervision, we augment the objective with a semantic density reward, derived from the precision-recall alignment between generated reasoning traces and ground-truth solutions. As a result, our method can mine more fine-grained optimization signals from a group of rollouts, guiding the LLM to a better optima. Experimental results across challenging math reasoning and question-answering tasks demonstrates that LambdaPO improves performance compared to the baseline methods.

Nico Uhlemann, Yipeng Zhou, Tobias Simeon Mohr et al.

This paper explores pedestrian trajectory prediction in urban traffic while focusing on both model accuracy and real-world applicability. While promising approaches exist, they often revolve around pedestrian datasets excluding traffic-related information, or resemble architectures that are either not real-time capable or robust. To address these limitations, we first introduce a dedicated benchmark based on Argoverse 2, specifically targeting pedestrians in traffic environments. Following this, we present Snapshot, a modular, feed-forward neural network that outperforms the current state of the art, reducing the Average Displacement Error (ADE) by 8.8% while utilizing significantly less information. Despite its agent-centric encoding scheme, Snapshot demonstrates scalability, real-time performance, and robustness to varying motion histories. Moreover, by integrating Snapshot into a modular autonomous driving software stack, we showcase its real-world applicability.

Wan Jiang, Gang Liu, Xiaofeng Chen et al.

Unlike traditional distributed machine learning, federated learning stores data locally for training and then aggregates the models on the server, which solves the data security problem that may arise in traditional distributed machine learning. However, during the training process, the transmission of model parameters can impose a significant load on the network bandwidth. It has been pointed out that the vast majority of model parameters are redundant during model parameter transmission. In this paper, we explore the data distribution law of selected partial model parameters on this basis, and propose a deep hierarchical quantization compression algorithm, which further compresses the model and reduces the network load brought by data transmission through the hierarchical quantization of model parameters. And we adopt a dynamic sampling strategy for the selection of clients to accelerate the convergence of the model. Experimental results on different public datasets demonstrate the effectiveness of our algorithm.

Jiwei Guan, Xi Zheng, Chen Wang et al.

With recent advances in autonomous driving, Voice Control Systems have become increasingly adopted as human-vehicle interaction methods. This technology enables drivers to use voice commands to control the vehicle and will be soon available in Advanced Driver Assistance Systems (ADAS). Prior work has shown that Siri, Alexa and Cortana, are highly vulnerable to inaudible command attacks. This could be extended to ADAS in real-world applications and such inaudible command threat is difficult to detect due to microphone nonlinearities. In this paper, we aim to develop a more practical solution by using camera views to defend against inaudible command attacks where ADAS are capable of detecting their environment via multi-sensors. To this end, we propose a novel multimodal deep learning classification system to defend against inaudible command attacks. Our experimental results confirm the feasibility of the proposed defense methods and the best classification accuracy reaches 89.2%. Code is available at https://github.com/ITSEG-MQ/Sensor-Fusion-Against-VoiceCommand-Attacks.

MiniMax, Aonian Li, Bangwei Gong et al.

We introduce MiniMax-01 series, including MiniMax-Text-01 and MiniMax-VL-01, which are comparable to top-tier models while offering superior capabilities in processing longer contexts. The core lies in lightning attention and its efficient scaling. To maximize computational capacity, we integrate it with Mixture of Experts (MoE), creating a model with 32 experts and 456 billion total parameters, of which 45.9 billion are activated for each token. We develop an optimized parallel strategy and highly efficient computation-communication overlap techniques for MoE and lightning attention. This approach enables us to conduct efficient training and inference on models with hundreds of billions of parameters across contexts spanning millions of tokens. The context window of MiniMax-Text-01 can reach up to 1 million tokens during training and extrapolate to 4 million tokens during inference at an affordable cost. Our vision-language model, MiniMax-VL-01 is built through continued training with 512 billion vision-language tokens. Experiments on both standard and in-house benchmarks show that our models match the performance of state-of-the-art models like GPT-4o and Claude-3.5-Sonnet while offering 20-32 times longer context window. We publicly release MiniMax-01 at https://github.com/MiniMax-AI.

Huitong Jin, Yipeng Zhou, Quan Z. Sheng et al.

Differentially Private Federated Learning (DPFL) strengthens privacy protection by perturbing model gradients with noise, though at the cost of reduced accuracy. Although prior empirical studies indicate that initializing from pre-trained rather than random parameters can alleviate noise disturbance, the problem of optimally fine-tuning pre-trained models in DPFL remains unaddressed. In this paper, we propose Pretrain-DPFL, a framework that systematically evaluates three most representative fine-tuning strategies: full-tuning (FT), head-tuning (HT), and unified-tuning(UT) combining HT followed by FT. Through convergence analysis under smooth non-convex loss, we establish theoretical conditions for identifying the optimal fine-tuning strategy in Pretrain-DPFL, thereby maximizing the benefits of pre-trained models in mitigating noise disturbance. Extensive experiments across multiple datasets demonstrate Pretrain-DPFL's superiority, achieving $25.22\%$ higher accuracy than scratch training and outperforming the second-best baseline by $8.19\%$, significantly improving the privacy-utility trade-off in DPFL.

Liming Wu, Wenbing Huang, Rui Jiao et al.

Crystal Structure Prediction (CSP), which aims to generate stable crystal structures from compositions, represents a critical pathway for discovering novel materials. While structure prediction tasks in other domains, such as proteins, have seen remarkable progress, CSP remains a relatively underexplored area due to the more complex geometries inherent in crystal structures. In this paper, we propose Siamese foundation models specifically designed to address CSP. Our pretrain-finetune framework, named DAO, comprises two complementary foundation models: DAO-G for structure generation and DAO-P for energy prediction. Experiments on CSP benchmarks (MP-20 and MPTS-52) demonstrate that our DAO-G significantly surpasses state-of-the-art (SOTA) methods across all metrics. Extensive ablation studies further confirm that DAO-G excels in generating diverse polymorphic structures, and the dataset relaxation and energy guidance provided by DAO-P are essential for enhancing DAO-G's performance. When applied to three real-world superconductors ($\text{CsV}_3\text{Sb}_5$, $ \text{Zr}_{16}\text{Rh}_8\text{O}_4$ and $\text{Zr}_{16}\text{Pd}_8\text{O}_4$) that are known to be challenging to analyze, our foundation models achieve accurate critical temperature predictions and structure generations. For instance, on $\text{CsV}_3\text{Sb}_5$, DAO-G generates a structure close to the experimental one with an RMSE of 0.0085; DAO-P predicts the $T_c$ value with high accuracy (2.26 K vs. the ground-truth value of 2.30 K). In contrast, conventional DFT calculators like Quantum Espresso only successfully derive the structure of the first superconductor within an acceptable time, while the RMSE is nearly 8 times larger, and the computation speed is more than 1000 times slower. These compelling results collectively highlight the potential of our approach for advancing materials science research and development.

Xiaoxin Su, Yipeng Zhou, Laizhong Cui et al.

Recently, federated learning (FL) has gained momentum because of its capability in preserving data privacy. To conduct model training by FL, multiple clients exchange model updates with a parameter server via Internet. To accelerate the communication speed, it has been explored to deploy a programmable switch (PS) in lieu of the parameter server to coordinate clients. The challenge to deploy the PS in FL lies in its scarce memory space, prohibiting running memory consuming aggregation algorithms on the PS. To overcome this challenge, we propose Federated Learning in-network Aggregation with Compression (FediAC) algorithm, consisting of two phases: client voting and model aggregating. In the former phase, clients report their significant model update indices to the PS to estimate global significant model updates. In the latter phase, clients upload global significant model updates to the PS for aggregation. FediAC consumes much less memory space and communication traffic than existing works because the first phase can guarantee consensus compression across clients. The PS easily aligns model update indices to swiftly complete aggregation in the second phase. Finally, we conduct extensive experiments by using public datasets to demonstrate that FediAC remarkably surpasses the state-of-the-art baselines in terms of model accuracy and communication traffic.

Xiaoxin Su, Yipeng Zhou, Laizhong Cui et al.

In Federated Learning (FL) paradigm, a parameter server (PS) concurrently communicates with distributed participating clients for model collection, update aggregation, and model distribution over multiple rounds, without touching private data owned by individual clients. FL is appealing in preserving data privacy; yet the communication between the PS and scattered clients can be a severe bottleneck. Model compression algorithms, such as quantization and sparsification, have been suggested but they generally assume a fixed code length, which does not reflect the heterogeneity and variability of model updates. In this paper, through both analysis and experiments, we show strong evidences that variable-length is beneficial for compression in FL. We accordingly present Fed-CVLC (Federated Learning Compression with Variable-Length Codes), which fine-tunes the code length in response of the dynamics of model updates. We develop optimal tuning strategy that minimizes the loss function (equivalent to maximizing the model utility) subject to the budget for communication. We further demonstrate that Fed-CVLC is indeed a general compression design that bridges quantization and sparsification, with greater flexibility. Extensive experiments have been conducted with public datasets to demonstrate that Fed-CVLC remarkably outperforms state-of-the-art baselines, improving model utility by 1.50%-5.44%, or shrinking communication traffic by 16.67%-41.61%.

Xianzhi Zhang, Yue Xu, Yinlin Zhu et al.

Multi-modal large language model (MLLM) inference scheduling enables strong response quality under practical and heterogeneous budgets, beyond what a homogeneous single-backend setting can offer. Yet online MLLM task scheduling is nontrivial, as requests vary sharply in modality composition and latent reasoning difficulty, while execution backends incur distinct, time-varying costs due to system jitter and network variation. These coupled uncertainties pose two core challenges: deriving semantically faithful yet scheduling-relevant multi-modal task representations, and making low-overhead online decisions over irreversible multi-dimensional budgets. Accordingly, we propose \emph{M-CMAB} (\underline{M}ulti-modal \underline{M}ulti-constraint \underline{C}ontextual \underline{M}ulti-\underline{A}rmed \underline{B}andit), a multi-adapter-enhanced MLLM inference scheduling framework with three components: (i) a CLS-attentive, frozen-backbone \emph{Predictor} that extracts compact task representations and updates only lightweight adapters for action-specific estimation; (ii) a primal-dual \emph{Constrainer} that maintains online Lagrange multipliers to enforce long-horizon constraints via per-round objectives; and (iii) a two-phase \emph{Scheduler} that balances exploration and exploitation under irreversible budgets. We establish a regret guarantee under multi-dimensional knapsack constraints. On a composite multimodal benchmark with heterogeneous backends, \emph{M-CMAB} consistently outperforms state-of-the-art baselines across budget regimes, achieving up to 14.18% higher reward and closely tracking an oracle-aided upper bound. Codes are available at https://anonymous.4open.science/r/M2CMAB/.

Ming Wen, Jiaqi Zhu, Yuedong Xu et al.

Large language models (LLMs) typically require fine-tuning for domain-specific tasks, and LoRA offers a computationally efficient approach by training low-rank adapters. LoRA is also communication-efficient for federated LLMs when multiple users collaboratively fine-tune a global LLM model without sharing their proprietary raw data. However, even the transmission of local adapters between a server and clients risks serious privacy leakage. Applying differential privacy (DP) to federated LoRA encounters a dilemma: adding noise to both adapters amplifies synthetic noise on the model, while fixing one adapter impairs the learnability of fine-tuning. In this paper, we propose FedASK (Differentially Private Federated Low Rank Adaptation with Double Sketching) , a novel federated LoRA framework to enable effective updating of both low-rank adapters with robust differential privacy. Inspired by randomized SVD, our key idea is a two-stage sketching pipeline. This pipeline first aggregates carefully sketched, privacy-preserving local updates, and then reconstructs the global matrices on the server to facilitate effective updating of both adapters. We theoretically prove FedASK's differential privacy guarantee and its exact aggregation property. Comprehensive experiments demonstrate that FedASK consistently outperforms baseline methods across a variety of privacy settings and data distributions.

Xianzhi Zhang, Yipeng Zhou, Miao Hu et al.

To mitigate the rising concern about privacy leakage, the federated recommender (FR) paradigm emerges, in which decentralized clients co-train the recommendation model without exposing their raw user-item rating data. The differentially private federated recommender (DPFR) further enhances FR by injecting differentially private (DP) noises into clients. Yet, current DPFRs, suffering from noise distortion, cannot achieve satisfactory accuracy. Various efforts have been dedicated to improving DPFRs by adaptively allocating the privacy budget over the learning process. However, due to the intricate relation between privacy budget allocation and model accuracy, existing works are still far from maximizing DPFR accuracy. To address this challenge, we develop BGTplanner (Budget Planner) to strategically allocate the privacy budget for each round of DPFR training, improving overall training performance. Specifically, we leverage the Gaussian process regression and historical information to predict the change in recommendation accuracy with a certain allocated privacy budget. Additionally, Contextual Multi-Armed Bandit (CMAB) is harnessed to make privacy budget allocation decisions by reconciling the current improvement and long-term privacy constraints. Our extensive experimental results on real datasets demonstrate that \emph{BGTplanner} achieves an average improvement of 6.76\% in training performance compared to state-of-the-art baselines.

Min Wen, Chengchang Liu, Ahmed Abdelmoniem et al.

Bilevel optimization, crucial for hyperparameter tuning, meta-learning and reinforcement learning, remains less explored in the decentralized learning paradigm, such as decentralized federated learning (DFL). Typically, decentralized bilevel methods rely on both gradients and Hessian matrices to approximate hypergradients of upper-level models. However, acquiring and sharing the second-order oracle is compute and communication intensive. % and sharing this information incurs heavy communication overhead. To overcome these challenges, this paper introduces a fully first-order decentralized method for decentralized Bilevel optimization, $\text{C}^2$DFB which is both compute- and communicate-efficient. In $\text{C}^2$DFB, each learning node optimizes a min-min-max problem to approximate hypergradient by exclusively using gradients information. To reduce the traffic load at the inner-loop of solving the lower-level problem, $\text{C}^2$DFB incorporates a lightweight communication protocol for efficiently transmitting compressed residuals of local parameters. % during the inner loops. Rigorous theoretical analysis ensures its convergence % of the algorithm, indicating a first-order oracle calls of $\tilde{\mathcal{O}}(ε^{-4})$. Experiments on hyperparameter tuning and hyper-representation tasks validate the superiority of $\text{C}^2$DFB across various typologies and heterogeneous data distributions.

Jiahao Tan, Yipeng Zhou, Gang Liu et al.

The federated learning (FL) paradigm emerges to preserve data privacy during model training by only exposing clients' model parameters rather than original data. One of the biggest challenges in FL lies in the non-IID (not identical and independently distributed) data (a.k.a., data heterogeneity) distributed on clients. To address this challenge, various personalized FL (pFL) methods are proposed such as similarity-based aggregation and model decoupling. The former one aggregates models from clients of a similar data distribution. The later one decouples a neural network (NN) model into a feature extractor and a classifier. Personalization is captured by classifiers which are obtained by local training. To advance pFL, we propose a novel pFedSim (pFL based on model similarity) algorithm in this work by combining these two kinds of methods. More specifically, we decouple a NN model into a personalized feature extractor, obtained by aggregating models from similar clients, and a classifier, which is obtained by local training and used to estimate client similarity. Compared with the state-of-the-art baselines, the advantages of pFedSim include: 1) significantly improved model accuracy; 2) low communication and computation overhead; 3) a low risk of privacy leakage; 4) no requirement for any external public information. To demonstrate the superiority of pFedSim, extensive experiments are conducted on real datasets. The results validate the superb performance of our algorithm which can significantly outperform baselines under various heterogeneous data settings.

Behnaz Soltani, Yipeng Zhou, Venus Haghighi et al.

In traditional machine learning, it is trivial to conduct model evaluation since all data samples are managed centrally by a server. However, model evaluation becomes a challenging problem in federated learning (FL), which is called federated evaluation in this work. This is because clients do not expose their original data to preserve data privacy. Federated evaluation plays a vital role in client selection, incentive mechanism design, malicious attack detection, etc. In this paper, we provide the first comprehensive survey of existing federated evaluation methods. Moreover, we explore various applications of federated evaluation for enhancing FL performance and finally present future research directions by envisioning some challenges.

Jiahao Liu, Jiang Wu, Jinyu Chen et al.

Different from conventional federated learning, personalized federated learning (PFL) is able to train a customized model for each individual client according to its unique requirement. The mainstream approach is to adopt a kind of weighted aggregation method to generate personalized models, in which weights are determined by the loss value or model parameters among different clients. However, such kinds of methods require clients to download others' models. It not only sheer increases communication traffic but also potentially infringes data privacy. In this paper, we propose a new PFL algorithm called \emph{FedDWA (Federated Learning with Dynamic Weight Adjustment)} to address the above problem, which leverages the parameter server (PS) to compute personalized aggregation weights based on collected models from clients. In this way, FedDWA can capture similarities between clients with much less communication overhead. More specifically, we formulate the PFL problem as an optimization problem by minimizing the distance between personalized models and guidance models, so as to customize aggregation weights for each client. Guidance models are obtained by the local one-step ahead adaptation on individual clients. Finally, we conduct extensive experiments using five real datasets and the results demonstrate that FedDWA can significantly reduce the communication traffic and achieve much higher model accuracy than the state-of-the-art approaches.

Yunchao Yang, Yipeng Zhou, Miao Hu et al.

Federated learning (FL) is a prospective distributed machine learning framework that can preserve data privacy. In particular, cross-silo FL can complete model training by making isolated data islands of different organizations collaborate with a parameter server (PS) via exchanging model parameters for multiple communication rounds. In cross-silo FL, an incentive mechanism is indispensable for motivating data owners to contribute their models to FL training. However, how to allocate the reward budget among different rounds is an essential but complicated problem largely overlooked by existing works. The challenge of this problem lies in the opaque feedback between reward budget allocation and model utility improvement of FL, making the optimal reward budget allocation complicated. To address this problem, we design an online reward budget allocation algorithm using Bayesian optimization named BARA (\underline{B}udget \underline{A}llocation for \underline{R}everse \underline{A}uction). Specifically, BARA can model the complicated relationship between reward budget allocation and final model accuracy in FL based on historical training records so that the reward budget allocated to each communication round is dynamically optimized so as to maximize the final model utility. We further incorporate the BARA algorithm into reverse auction-based incentive mechanisms to illustrate its effectiveness. Extensive experiments are conducted on real datasets to demonstrate that BARA significantly outperforms competitive baselines by improving model utility with the same amount of reward budget.

Laizhong Cui, Xiaoxin Su, Yipeng Zhou et al.

Federated Learning (FL) incurs high communication overhead, which can be greatly alleviated by compression for model updates. Yet the tradeoff between compression and model accuracy in the networked environment remains unclear and, for simplicity, most implementations adopt a fixed compression rate only. In this paper, we for the first time systematically examine this tradeoff, identifying the influence of the compression error on the final model accuracy with respect to the learning rate. Specifically, we factor the compression error of each global iteration into the convergence rate analysis under both strongly convex and non-convex loss functions. We then present an adaptation framework to maximize the final model accuracy by strategically adjusting the compression rate in each iteration. We have discussed the key implementation issues of our framework in practical networks with representative compression algorithms. Experiments over the popular MNIST and CIFAR-10 datasets confirm that our solution effectively reduces network traffic yet maintains high model accuracy in FL.

Yipeng Zhou, Xuezheng Liu, Yao Fu et al.

Federated learning (FL) empowers distributed clients to collaboratively train a shared machine learning model through exchanging parameter information. Despite the fact that FL can protect clients' raw data, malicious users can still crack original data with disclosed parameters. To amend this flaw, differential privacy (DP) is incorporated into FL clients to disturb original parameters, which however can significantly impair the accuracy of the trained model. In this work, we study a crucial question which has been vastly overlooked by existing works: what are the optimal numbers of queries and replies in FL with DP so that the final model accuracy is maximized. In FL, the parameter server (PS) needs to query participating clients for multiple global iterations to complete training. Each client responds a query from the PS by conducting a local iteration. Our work investigates how many times the PS should query clients and how many times each client should reply the PS. We investigate two most extensively used DP mechanisms (i.e., the Laplace mechanism and Gaussian mechanisms). Through conducting convergence rate analysis, we can determine the optimal numbers of queries and replies in FL with DP so that the final model accuracy can be maximized. Finally, extensive experiments are conducted with publicly available datasets: MNIST and FEMNIST, to verify our analysis and the results demonstrate that properly setting the numbers of queries and replies can significantly improve the final model accuracy in FL with DP.

Laizhong Cui, Xiaoxin Su, Yipeng Zhou et al.

Federated Learning (FL) is an emerging decentralized learning framework through which multiple clients can collaboratively train a learning model. However, a major obstacle that impedes the wide deployment of FL lies in massive communication traffic. To train high dimensional machine learning models (such as CNN models), heavy communication traffic can be incurred by exchanging model updates via the Internet between clients and the parameter server (PS), implying that the network resource can be easily exhausted. Compressing model updates is an effective way to reduce the traffic amount. However, a flexible unbiased compression algorithm applicable for both uplink and downlink compression in FL is still absent from existing works. In this work, we devise the Model Update Compression by Soft Clustering (MUCSC) algorithm to compress model updates transmitted between clients and the PS. In MUCSC, it is only necessary to transmit cluster centroids and the cluster ID of each model update. Moreover, we prove that: 1) The compressed model updates are unbiased estimation of their original values so that the convergence rate by transmitting compressed model updates is unchanged; 2) MUCSC can guarantee that the influence of the compression error on the model accuracy is minimized. Then, we further propose the boosted MUCSC (B-MUCSC) algorithm, a biased compression algorithm that can achieve an extremely high compression rate by grouping insignificant model updates into a super cluster. B-MUCSC is suitable for scenarios with very scarce network resource. Ultimately, we conduct extensive experiments with the CIFAR-10 and FEMNIST datasets to demonstrate that our algorithms can not only substantially reduce the volume of communication traffic in FL, but also improve the training efficiency in practical networks.

Xianzhi Zhang, Yipeng Zhou, Di Wu et al.

It is always a challenging problem to deliver a huge volume of videos over the Internet. To meet the high bandwidth and stringent playback demand, one feasible solution is to cache video contents on edge servers based on predicted video popularity. Traditional caching algorithms (e.g., LRU, LFU) are too simple to capture the dynamics of video popularity, especially long-tailed videos. Recent learning-driven caching algorithms (e.g., DeepCache) show promising performance, however, such black-box approaches are lack of explainability and interpretability. Moreover, the parameter tuning requires a large number of historical records, which are difficult to obtain for videos with low popularity. In this paper, we optimize video caching at the edge using a white-box approach, which is highly efficient and also completely explainable. To accurately capture the evolution of video popularity, we develop a mathematical model called \emph{HRS} model, which is the combination of multiple point processes, including Hawkes' self-exciting, reactive and self-correcting processes. The key advantage of the HRS model is its explainability, and much less number of model parameters. In addition, all its model parameters can be learned automatically through maximizing the Log-likelihood function constructed by past video request events. Next, we further design an online HRS-based video caching algorithm. To verify its effectiveness, we conduct a series of experiments using real video traces collected from Tencent Video, one of the largest online video providers in China. Experiment results demonstrate that our proposed algorithm outperforms the state-of-the-art algorithms, with 12.3\% improvement on average in terms of cache hit rate under realistic settings.

Yao Fu, Yipeng Zhou, Di Wu et al.

In spite that Federated Learning (FL) is well known for its privacy protection when training machine learning models among distributed clients collaboratively, recent studies have pointed out that the naive FL is susceptible to gradient leakage attacks. In the meanwhile, Differential Privacy (DP) emerges as a promising countermeasure to defend against gradient leakage attacks. However, the adoption of DP by clients in FL may significantly jeopardize the model accuracy. It is still an open problem to understand the practicality of DP from a theoretic perspective. In this paper, we make the first attempt to understand the practicality of DP in FL through tuning the number of conducted iterations. Based on the FedAvg algorithm, we formally derive the convergence rate with DP noises in FL. Then, we theoretically derive: 1) the conditions for the DP based FedAvg to converge as the number of global iterations (GI) approaches infinity; 2) the method to set the number of local iterations (LI) to minimize the negative influence of DP noises. By further substituting the Laplace and Gaussian mechanisms into the derived convergence rate respectively, we show that: 3) The DP based FedAvg with the Laplace mechanism cannot converge, but the divergence rate can be effectively prohibited by setting the number of LIs with our method; 4) The learning error of the DP based FedAvg with the Gaussian mechanism can converge to a constant number finally if we use a fixed number of LIs per GI. To verify our theoretical findings, we conduct extensive experiments using two real-world datasets. The results not only validate our analysis results, but also provide useful guidelines on how to optimize model accuracy when incorporating DP into FL

Yupeng Jiang, Yong Li, Yipeng Zhou et al.

In federated learning, machine learning and deep learning models are trained globally on distributed devices. The state-of-the-art privacy-preserving technique in the context of federated learning is user-level differential privacy. However, such a mechanism is vulnerable to some specific model poisoning attacks such as Sybil attacks. A malicious adversary could create multiple fake clients or collude compromised devices in Sybil attacks to mount direct model updates manipulation. Recent works on novel defense against model poisoning attacks are difficult to detect Sybil attacks when differential privacy is utilized, as it masks clients' model updates with perturbation. In this work, we implement the first Sybil attacks on differential privacy based federated learning architectures and show their impacts on model convergence. We randomly compromise some clients by manipulating different noise levels reflected by the local privacy budget epsilon of differential privacy on the local model updates of these Sybil clients such that the global model convergence rates decrease or even leads to divergence. We apply our attacks to two recent aggregation defense mechanisms, called Krum and Trimmed Mean. Our evaluation results on the MNIST and CIFAR-10 datasets show that our attacks effectively slow down the convergence of the global models. We then propose a method to keep monitoring the average loss of all participants in each round for convergence anomaly detection and defend our Sybil attacks based on the prediction cost reported from each client. Our empirical study demonstrates that our defense approach effectively mitigates the impact of our Sybil attacks on model convergence.

Jiqiang Wu, Yipeng Zhou, Dah Ming Chiu et al.

Large Internet video delivery systems serve millions of videos to tens of millions of users on daily basis, via Video-on-Demand and live streaming. Video popularity evolves over time. It represents the workload, as welll as business value, of the video to the overall system. The ability to predict video popularity is very helpful for improving service quality and operating efficiency. Previous studies adopted simple models for video popularity, or directly adopted patterns from measurement studies. In this paper, we develop a stochastic fluid model that tries to capture two hidden processes that give rise to different patterns of a given video's popularity evolution: the information spreading process, and the user reaction process. Specifically, these processes model how the video is recommended to the user, the videos inherent attractiveness, and users reaction rate, and yield specific popularity evolution patterns. We then validate our model by matching the predictions of the model with observed patterns from our collaborator, a large content provider in China. This model thus gives us the insight to explain the common and different video popularity evolution patterns and why.

Liang Chen, Yipeng Zhou, Dah Ming Chiu

Online video-on-demand(VoD) services invariably maintain a view count for each video they serve, and it has become an important currency for various stakeholders, from viewers, to content owners, advertizers, and the online service providers themselves. There is often significant financial incentive to use a robot (or a botnet) to artificially create fake views. How can we detect the fake views? Can we detect them (and stop them) using online algorithms as they occur? What is the extent of fake views with current VoD service providers? These are the questions we study in the paper. We develop some algorithms and show that they are quite effective for this problem.

Liang Chen, Yipeng Zhou, Dah Ming Chiu

Bandwidth consumption is a significant concern for online video service providers. Practical video streaming systems usually use some form of HTTP streaming (progressive download) to let users download the video at a faster rate than the video bitrate. Since users may quit before viewing the complete video, however, much of the downloaded video will be "wasted". To the extent that users' departure behavior can be predicted, we develop smart streaming that can be used to improve user QoE with limited server bandwidth or save bandwidth cost with unlimited server bandwidth. Through measurement, we extract certain user behavior properties for implementing such smart streaming, and demonstrate its advantage using prototype implementation as well as simulations.