Ying Liu

Video anomaly detection (VAD) is important in many fields because of its theoretical and practical values. One of the challenges in VAD is the difficulty in obtaining segment-level labels due to the high annotation cost. In recent years, researchers have adopted video-level labels as a form of weak supervision, leading to the development of weakly supervised video anomaly detection (WS-VAD). Among different WS-VAD approaches, graph convolutional networks (GCNs) have attracted much attention, since they have the ability to model relationship information in video data. Typically, the relationship, represented by the graph edges, is the class label similarity, and this similarity is built based on the feature similarity and temporal consistency among video segments. Undoubtedly, the more information about class label similarity is provided, the higher the performance of GCN tends to be. In real-world scenarios of VAD, anomalies exhibit several unique properties such as diversity and rarity. These properties may lead to the following situation. Given two video segments, although their feature similarity is low and their time separation is large, both of them are anomalies, that is, they have the same class label. Likewise, normal samples also encounter such situation. However, the existing graph structures in GCN methods do not adequately account for this situation. To address this issue, this paper proposes an extended graph learning (EGL) method that incorporates additional class label similarity among video segments. The proposed EGL includes two extended graph convolutional networks (EGCNs): a spatial EGCN and a temporal EGCN. To capture more accurate information about class label similarity, EGL incorporates a feedback module to update the graph structures of EGCNs. EGL can effectively extract more information about class label similarity, thereby ensuring good performance when training data is scarce. Experimental results highlight the advantages of the proposed EGL method, particularly with limited training samples. In particular, when only 30% of the training data is used, EGL achieves the best performance of 95.55% AUC on ShanghaiTech, 81.29% AUC on UCF-Crime, and 75.31% AP on XD-Violence, outperforming the existing VAD methods by up to 6.86%, 3.23%, and 4.40%, respectively.

Federated learning (FL), which enables collaborative learning across distributed nodes, confronts a significant heterogeneity challenge, primarily including resource heterogeneity induced by different hardware platforms, and statistical heterogeneity originating from non-IID private data distributions among clients. Neural architecture search (NAS), particularly one-shot NAS, holds great promise for automatically designing optimal personalized models tailored to such heterogeneous scenarios. However, the coexistence of both resource and statistical heterogeneity destabilizes the training of the one-shot supernet, impairs the evaluation of candidate architectures, and ultimately hinders the discovery of optimal personalized models. To address this problem, we propose a heterogeneity-aware personalized federated NAS (HAPFNAS) method. First, we leverage lightweight knowledge models to distill knowledge from clients to server-side supernet, thereby effectively mitigating the effects of heterogeneity and enhancing the training stability. Then, we build random-forest-based personalized performance predictors to enable the efficient evaluation of candidate architectures across clients. Furthermore, we develop a model-heterogeneous FL algorithm called heteroFedAvg to facilitate collaborative model training for the discovered personalized models. Comprehensive experiments on CIFAR-10/100 and Tiny-ImageNet classification datasets demonstrate the effectiveness of our HAPFNAS, compared to state-of-the-art federated NAS methods.

Vertical Federated Learning (VFL), which enables multiple parties that hold different features of the same data set to collaboratively train a global learning model without sharing their raw data, is an important paradigm in FL. However, VFL is vulnerable to privacy threats, particularly Label Inference Attacks (LIA), where passive participants try to deduce sensitive labels of the active party. Existing defense strategies, such as Homomorphic Encryption (HE) and Differential Privacy (DP), often face challenges related to high computational overhead or tradeoff between privacy and accuracy, especially in asynchronous settings. Considering this, in this paper, we propose a novel defense mechanism against batch-level label inference attacks (BLI) in vertical asynchronous federated learning (VAFL), where client-level differential noise is added to obscure batch-level gradients such that the passive participants cannot derive the label information from the active participant. Besides, to address party's asynchronicity, our approach incorporates a temporal weight decay strategy according to the staleness of each party. A series numerical simulations are then performed to verify the effectiveness and robustness in privacy protection for VAFL scenarios.

Distributed multitask learning, which aims to estimate multiple parameter vectors in a distributed and collaborative manner, has been applied to a wide range of Internet of Things (IoTs). However, in many practical applications, such as channel impulse responses (CIRs) estimation and adaptive line enhancer (ALE), the parameter vectors of interest can be of large size, resulting in high computational cost. Frequency-domain (FD) learning that utilizes efficient Fast Fourier Transform (FFT) may be a good candidate to reduce computational complexity. Moreover, benefited from the pre-whitening effect caused by the orthogonality property of FFT, the estimation performance can be improved. However, existing research on FD learning mainly focuses on single-task scenarios, which is not applicable to multitask scenarios. Considering this, in this article, we first formulate the FD distributed multitask learning problem, and then propose a frequency-domain distributed multitask learning algorithm. Through performing numerical simulations and practical applications to CIRs estimation and smart grid state estimation, it is observed that the proposed algorithm exhibits lower computational complexity and better estimation performance, compared with the existing time-domain distributed algorithms.

Neural architecture search (NAS) has shown great promise in automatically designing neural network models. Recently, block-wise NAS has been proposed to alleviate deep coupling problem between architectures and weights existed in the well-known weight-sharing NAS, by training the huge weight-sharing supernet block-wisely. However, the existing block-wise NAS methods, which resort to either supervised distillation or self-supervised contrastive learning scheme to enable block-wise optimization, take massive computational cost. To be specific, the former introduces an external high-capacity teacher model, while the latter involves supernet-scale momentum model and requires a long training schedule. Considering this, in this work, we propose a resource-friendly deeply supervised block-wise NAS (DBNAS) method. In the proposed DBNAS, we construct a lightweight deeply-supervised module after each block to enable a simple supervised learning scheme and leverage ground-truth labels to indirectly supervise optimization of each block progressively. Besides, the deeply-supervised module is specifically designed as structural and functional condensation of the supernet, which establishes global awareness for progressive block-wise optimization and helps search for promising architectures. Experimental results show that the DBNAS method only takes less than 1 GPU day to search out promising architectures on the ImageNet dataset with less GPU memory footprint than the other block-wise NAS works. The best-performing model among the searched DBNAS family achieves 75.6% Top-1 accuracy on ImageNet, which is competitive with the state-of-the-art NAS models. Moreover, our DBNAS family models also achieve good transfer performance on CIFAR-10/100, as well as two downstream tasks: object detection and semantic segmentation.

Distributed classification, as an important distributed learning task, has received much attention for more than two decades. However, it is still a challenge to achieve acceptable classification performance on some real datasets that follow complex distributions. Moreover, in various practical scenarios, some data features are unavoidably missing due to environmental interference or human factors, which affect the performance of classification. Considering this, in this paper, we propose a distributed Gaussian Mixture Model-based Multi-Kernel Learning (dGMM-MKL) algorithm for classification of missing data distributed over a network. In the proposed algorithm, a distributed imputation method based on Gaussian Mixture Model (GMM) is used to make imputation for the missing components. Then, a distributed multi-kernel support vector machine (SVM) is developed to achieve classification based on the imputed data. Finally, we present experiment results on some datasets. Compared with state-of-the-art algorithms, the proposed dGMM-MKL algorithm shows better classification performance and higher robustness under different missing probabilities.

Fast and efficient tobacco tax profit predication is essential to corporate budget computation and business decision-making, which provides an important guarantee for the development of tobacco industry. In this paper, we first use Pearson correlation coefficient analysis and Lasso sparse model to analyze the effect of the external factors including macroeconomic indicators, consumer characteristic factors, and population index on the tobacco tax profit, and excavate those external factors that significantly affect the tobacco tax profit. After that, based on the selected external factors and also considering the influence of tax policy of the government on tobacco industry, a policy-based hybrid Holt-SVR model is proposed to obtain the interval predication of the tobacco tax profit. Experimental results show that, the proposed interval prediction method based on factor analysis improves the performance of prediction, compared with the traditional multiple regression models. Besides, the usage of policy variable for constructing the predication model also effectively improves the robustness of the proposed method.

Federated learning (FL) has attracted widespread attention in the Internet of Things domain recently. With FL, multiple distributed devices can cooperatively train a global model by transmitting model updates without disclosing the original data. However, the distributed nature of FL makes it vulnerable to data poisoning attacks. In practice, malicious clients can launch the label-flipping attack (LFA) by simply tampering with the labels of local data, thus causing the global model to misclassify the samples of a selected class as the target class. Although some defense mechanisms have been proposed, they rely on specific assumptions about data distribution, and their performance degrades significantly when the data on clients are non-IID. Besides, most existing methods require clients to upload model updates in plaintext so that the server can identify and remove the malicious updates. But, direct transmission of model updates may still reveal private information. Considering these issues, we develop a label-flipping-robust and privacy-preserving FL (LFR-PPFL) algorithm, which is applicable to both independent and identically distributed (IID) and non-IID data. We first propose a detection method based on temporal analysis on cosine similarity to distinguish malicious clients from benign clients. Then, we propose a privacy-preserving computation protocol based on homomorphic encryption to implement this detection method and perform federated aggregation while protecting the privacy of clients. Besides, a detailed theoretical analysis is given to demonstrate the privacy guarantee of the proposed protocol. Experimental results on real-world data sets show that the proposed algorithm can effectively defend against LFAs under various data distributions.