With the rapid advancement of cloud computing technologies, microservice architectures have emerged as the de facto standard for modern software systems. However, their highly dynamic and complex nature poses significant challenges for anomaly detection. Although recent diffusion-based models have demonstrated promising capabilities in modeling multi-variate time series, they often overlook the inherent multi-scale spatiotemporal correlations present in microservice systems. This oversight limits their capacity to fully capture system-wide behaviors, thereby constraining detection performance. To address this issue, we propose a novel framework for anomaly detection based on imputed diffusion models, termed Multi-Scale Graph Imputed Diffusion Model. Building upon the imputation–reconstruction paradigm of standard ImDiffusion, our approach introduces a multi-scale spatiotemporal correlation module that is seamlessly integrated into the denoising network. This module employs parallel multi-scale graph convolutions and temporal convolutions to explicitly capture inter-service spatial dependencies from the microservice dependency graph and multi-granular temporal dynamics from service metric time series. By conditioning the diffusion process on richer spatiotemporal contexts, MSG-ImDiffusion enhances the model’s sensitivity to anomalies that deviate from expected temporal and spatial patterns. Extensive experiments conducted on the open-source HipsterShop microservice benchmark demonstrate that our proposed model significantly outperforms existing baselines, including ImDiffusion, in terms of both detection accuracy and F1-score. These results validate the effectiveness of incorporating multi-scale spatiotemporal information into diffusion-based anomaly detection frameworks for microservice systems.

Microservice architecture has been widely adopted in large-scale applications. However, it also brings new challenges to ensuring reliable performance and maintenance due to the huge volume of data and complex dependencies of microservices. Existing approaches still suffer from the over-aggregation of data, interference from anomaly propagation, and ignoration of component differences. To solve these issues, this paper builds a root cause diagnosis framework at the operation granularity, named as MicroNet. Since operations are subfunctions of microservices, recorded as invocation purposes, we propose the operation-centric perspective, to realize fine-grained data aggregation and operation-level anomaly backtracking. We decompose the diagnosis task into four phases: dependency graph construction, anomaly detection, anomaly evaluation, and culprit location. To construct the invocation dependency accurately, we propose the concept of meta call, defined as the triple (caller, operation, callee), the smallest unit that can be aggregated. Based on the dependency graph, we quantify the operation’s abnormality by analyzing the operation execution process, to backtrack the propagated anomalies. Then, we customize a personalized PageRank algorithm to identify the root cause in which invocation latency and different invocation relationships are considered simultaneously. Our experimental evaluation on an open dataset shows that MicroNet can effectively locate root causes with 90% mean average precision, outperforming state-of-the-art methods.

In a medical dialogue diagnosis system, the selection of symptoms for inquiry has a significant impact on diagnostic accuracy and dialogue efficiency. In a typical diagnosis process, the symptoms initially reported by users are often insufficient to support an accurate diagnosis, making it necessary to ask users about other symptoms through dialogue to form a conclusive diagnosis. In this paper, we propose a disease diagnosis algorithm based on Bayesian, which simulates the process of doctor's inquiry and diagnosis by dynamically updating the list of diseases to increase the interpretability of diagnosis results. For the symptom interrogation, we propose a symptom screening algorithm based on the difference of symptom sets to exclude diseases with low probability. Through the intersection and union of disease symptom sets, we can screen out the symptoms that can distinguish diseases in fewer inquiring rounds. The experimental results demonstrate the proposed method performs more efficiently than existing state-of-the-art algorithms.

Microservice architecture has been widely adopted for large-scale applications because of its benefits of scalability, flexibility, and reliability. However, due to the heterogeneity of system architecture and variable invocations between services, it is difficult to accurately diagnose the root causes of performance degradation in time. This paper proposes WinG, a system to pinpoint root causes. Firstly, since the characteristic of a system element is difficult to capture due to variable invocations between elements, WinG characterizes an element's status by a feature vector that includes all its invocation relationships. Secondly, WinG adopts a warping procedure to assess an element's anomaly severity based on its status deviation, to mitigate the interference of variable invocations. Thirdly, WinG groups heterogeneous elements with similar invocation characteristics to avoid the interference of diverse elements types. Finally, false alarms are filtered by the anomaly duration and frequency. Experimental evaluation results on the public dataset show that, with the above four methods, WinG can locate root causes with 87% precision, outperforming baseline methods. On average of 78 test cases, it achieves 34% precision improvement over the champion method of the competition.

As an application of the Internet of Things in the automotive field, Vehicular Ad-hoc NETworks (VANETs) are developed to facilitate traffic safety and traffic flow optimization. VANET consists of the communication network and the underlay road network. Due to the characteristics of open access, the communication network is vulnerable to various attacks, especially, the information pollution attack which is highly risky and concealed. Such an attack can lead the vehicles to react to the false messages from attackers, even cause failure cascades. However, the impact of the information pollution attack on the road network has not gained enough attention. To assess such impact, we build a traffic model with polluted information and assess the vulnerability of the road network under different pollution scenarios in terms of attack percentage, attack types, and road network topologies. Experimental results demonstrate that VANETs are vulnerable to the information pollution attack, and transportation performance drops by half when only 15% of edges are attacked in the worst case.

With the development of the smart city, personalized path recommendation has already attracted the attention of researchers. However, it has not been considered that some users need to make personalized recommendations of a route for a given pair of OD (Origin-Destination) and some specified consecutive way-points. To our knowledge, this problem is studied for the first time. Essentially, this problem can be taken as inferring a high sampling rate fine trajectory from a low sampling rate rough trajectory composed of OD and way-points. The biggest challenge is that the user may just have some knowledge about the location rather than the precise location of a waypoint because a place may have many GPS points. This paper proposes a PSR (Personalized Selective Route) model based on trajectory learning for a given OD and a number of way-points. It integrates multi-source information to learn more comprehensive personalized preferences and introduces the Seq2Seq model with Multi-Head self-attention mechanism to automatically adjust the weights to capture more accurate temporal and spatial correlations. The experiments on the real traffic trajectory data show PSR model is robust to low sampling rate and noise. Compared with the best baseline, the accuracy of Top-l under the Euclidean distance of PSR is improved by 32.57%, and the accuracy of Top-3 is improved by 63.87%.

Nowadays, a lot of data mining applications, such as web traffic analysis and content popularity prediction, leverage users’ web browsing trajectories to improve their performance. However, the disclosure of web browsing trajectory is the most prominent issue. A novel privacy model, named Differential Privacy, is used to rigorously protect user’s privacy. Some works have applied this privacy model to spatial-temporal streams. However, these works either protect the users’ activities in different places separately or protect their activities in all places jointly. The former one cannot protect trajectories that traverse multiple places; while the latter ignores the differences among places and suffers the degradation of data utility (i.e., data accuracy). In this paper, we propose a w , n -differential privacy to protect any spatial-temporal sequence occurring in w successive timestamps and n -range places. To achieve better data utility, we propose two implementation algorithms, named Spatial-Temporal Budget Distribution (STBD) and Spatial-Temporal RescueDP (STR). Theoretical analysis and experimental results show that these two algorithms can achieve a balance between data utility and trajectory privacy guarantee.

With the boosting of application, biometric identification with fingerprint or face-image suffers from forging attacks. The electrocardiogram (ECG) as a kind of biometric identification is of higher resistance against such attacks and receives research attention. The state-of-art method has recognition accuracy of about 95%. However, we find that the accuracy will degrade dramatically to 40% if it is applied in a practical context when a significant interval between training period and applying period. The critical reasons for this failure are as follows: 1) the extracted features are temporal sensitive due to that continuous samples being used in training and testing period in the existing schemes; 2) the features highly relevant to the performance are not utilized sufficiently in CNN classifier; 3) the optimal parameter setting for obtaining enough effective samples for individuals has not been investigated. This paper targets on proposing a practical cross- domain ECG biometric identification method to solve the above problems.Specifically,we:1)determine the best parameters of the non-fiducial random sampling method to obtain enough effective samples for individuals; 2) propose a method to extract deep features across time, frequency and energy domain which are temporal insensitive and individual distinguishable; 3) introduce a channel attention module into the CNN and modify its activation function to optimize the recognition performance. We validate our method on PTBDB and ECG-ID databases. Experiments show that the identification accuracy reaches 56.93% and 85.94% respectively, with an improvement of 41.5% and 20.7% over the existing method.

Continuously sharing user's trajectory data which contain one's location information makes the crowd sensing of the traffic dynamics and mobility trends feasible. This kind of spatial streaming data is beneficial for intelligent transportation but at the risk of disclosing personal privacy, even if it is published in statistical form such as “the number of users in an area at time t”. The user number on a location at time t is similar to that of previous release on the same location, and to that on adjacent locations. Such spatio-temporal correlation makes it a challenge to find solutions to protect user's trajectory privacy. The state-of-the-art privacy protection framework, differential privacy, has been extended to streaming scenario for preventing the privacy leak causing by the temporal correlation. However, such schemes neglect the importance of spatial correlation so that they may suffer the leak of user trajectory privacy or the degradation of data utility. Based on the observation that any piece of trajectory has temporal and spatial locality, we propose a flexible trajectory privacy model of w-event n2-block differential privacy, short as (ω, n)-differential privacy, to ensure any trajectory occurring in an area of n×n blocks during w successive timestamps under the protection of ε-differential privacy. Then we design the Spatial Temporal Budget Distribution (STBD) algorithm for achieving (ω, n)-differential privacy. Validation results of this algorithm on two real-life datasets and one synthetic dataset confirm its practicality.

With the surging demand on high‐quality mobile video services and the unabated development of new network technology, including fog computing, there is a need for a generalized quality of user experience (QoE) model that could provide insight for various network optimization designs. A good QoE, especially when measured as engagement, is an important optimization goal for investors and advertisers. Therefore, many works have focused on understanding how the factors, especially quality of service (QoS) factors, impact user engagement. However, the divergence of user interest is usually ignored or deliberatively decoupled from QoS and/or other objective factors. With an increasing trend towards personalization applications, it is necessary as well as feasible to consider user interest to satisfy aesthetic and personal needs of users when optimizing user engagement. We first propose anExtraction-Inference (E-I)algorithm to estimate the user interest from easily obtained user behaviors. Based on our empirical analysis on a large‐scale dataset, we then build aQoS and user Interest based Engagement (QI-E) regression model. Through experiments on our dataset, we demonstrate that the proposed model reaches an improvement in accuracy by 9.99% over the baseline model which only considers QoS factors. The proposed model has potential for designing QoE‐oriented scheduling strategies in various network scenarios, especially in the fog computing context.

With the prevailing of mobile phones and online video contents, the demand for mobile online video is increasing. The desire, however, is held down by the high mobile traffic cost. To solve this problem, an off-the-shelf solution is WiFi-Direct (WFD), which is widely available on a majority of mobile devices. There is, however, no systematic study of WFD-based group data transfer and cooperative video streaming on real phones. Thus, we designed and implemented a WFD-based LTE cooperative video streaming system, in which the WFD GO (Group Owner) device takes the responsibility of peer information exchange, data relay, and LTE cooperative downloading scheduling. Based on the system, we evaluated the performance of WFD-based group data sharing, including Ping response delay, throughput, and power efficiency. Valuable findings were obtained. For instance, we discovered that when a WFD device connects to a traditional AP (Access Point), even if there is \emph{no} data transmission to/from the AP, the Device-to-Device (D2D) throughput would decrease by at least 72\%. Based on these findings, we provided recommendations for the design and deployment of WFD based D2D systems. We finally demonstrated the feasibility of WFD based LTE cooperative video streaming using our system. We showed that, using multiple realistic LTE networks, a 3-device cooperative system can provide smooth video streaming with bitrate more than 10Mbps.

Some recent work proposed differential private collaborative filtering (DPCF) recommender systems to protect user privacy from indirect access attacks, e.g., KNN attacks. As the cost of such protection, the MAE of recommendation was retained with insignificant increase but the more focused metrics, i.e., the precision and recall of top-k recommendations, degraded unacceptable. To address this problem, we propose a DPCF framework based on privacy-relevance of topics, named DPCFT. Firstly, DPCFT works on topic-preference level which highly aggregates user behaviors and keeps the precision and recall of top-k differential private recommendations acceptable. More importantly, considering the unnoticed fact that the information leakage of some special topics worries users much more than that of other ones in terms of privacy concerns, DPCFT introduces the topic privacy-relevance level in the similarity computation and neighbor selection to impose stronger privacy protection on higher privacy-relevance topics with overall differential privacy and recommendation performance reserved. Finally, to reduce the recommendation performance cost for differential privacy, DPCFT selects the top-k recommendation items at user side further with personal topic-preference without risk of data expose. Experimental results on the MovieLens dataset verify that the proposed framework DPCFT preserves differential privacy and top-k recommendation performance simultaneously.

Video watching time is a crucial measure for studying user watching behavior in online Internet video-on-demand (VoD) systems. It is important for system planning, user engagement study, and service quality evaluation. However, due to limited access to large-scale VoD systems, there is still a lack of accurate model for characterizing the distribution of user watching time on a per video basis. In this paper, we measure PPLive, one of the most popular commercial Internet VoD systems in China, over a three week period, and characterize user watching time distributions of 1,000 most popular movies. We find that a video's watching time can be modeled by a concatenation of exponential distribution (in the first several minutes of the video) and truncated power law distribution (in the remaining time of the video), when users watch the video without interruptions. For comparison, user watching time with user interactions such as seeking and/or pause operations does not follow such a distribution. We further reveal interesting characteristics regarding the relation between video's watching time distribution and various watching/video-related features (including time-of-day, user ratings, and movie genres). Our measurement and modeling results bring forth important insights for design, deployment, and evaluation of Internet VoD systems.

In online video systems, viewer demographic information (gender, age, etc.) is of huge commercial value for delivering targeted advertising and video recommendations, but generally not available directly. This paper targets inferring viewers' gender based on implicit watching history in the large-scale online video systems. To tackle the sparsity problem without filtering out any cold users or videos, we not only introduce video tags as features, but also use an efficient Chinese word segmentation method to extract hot key-words from video titles as features. Moreover, users' viewing behavior distribute lognormally, hence we apply a logarithmic transformation on the inference matrixes and further find key features via principal components analysis (PCA). We then solve the gender inference as a classification problem and define some modified evaluation metrics adapt to the imbalance classification problem. We compare a set of classifiers including Class prior, EM, SVM, Logistic regression, Partially supervised soft-label and belief-based mixture and find that Logistic regression is the best. The inference results show that our algorithms can obtain high F̃1values for all classes. The highest value of PPTV dataset can reach nearly 0.75. And inference based on key-words results in a 14.63% increase of F̃1contrast to the ratings of MovieLens.

Based on a large-scale dataset extracted from the servers' logs of PPTV, one of the largest online video service providers in China, we study how device types, wireless network connection types and video qualities impact user's watching behaviors and network traffic. We found that 1) the diurnal user viewing patterns on mobile devices are slightly different from that on PC devices; 2) the mobile APP's landing page is the primary source for users to find videos to watch, and the keyword search is the secondary source; 3) with respect to video quality, iOS devices are better than Android devices and tablet are better than smart phones; and 4) comparing to users with PC devices, users using mobile devices watch shorter, but are more concentrated on popular videos so that the video popularity distribution is more skewed. We further provide insights and suggestions in providing mobile video services and improving services' quality.

A peer-to-peer (P2P) live streaming system faces a big challenge under flash crowds. When a flash crowd occurs, the sudden arrival of numerous peers may starve the upload capacity of the system, hurt its quality of service, and even cause system collapse. This paper provides a comprehensive study on the performance of P2P live streaming systems under flash crowds. By modeling the systems using a fluid model, we study the system capacity, peer startup latency, and system recovery time of systems with and without admission control for flash crowds, respectively. Our study demonstrates that, without admission control, a P2P live streaming system has limited capacity to handle flash crowds. We quantify this capacity by the largest flash crowd (measured in shock level) that the system can handle, and further find this capacity is independent of system initial state while decreasing as departure rate of stable peer increases, in a power-law relationship. We also establish the mathematical relationship of flash crowd size to the worst-case peer startup latency and system recovery time. For a system with admission control, we prove that it can recover stability under flash crowds of any sizes. Moreover, its worst-case peer startup latency and system recovery time increase logarithmically with the flash crowd size. Based on the analytical results, we present detailed flash crowd handling strategies, which can be used to achieve satisfying peer startup performance while keeping system stability in the presence of flash crowds under different circumstances .

Video watching time is a crucial measure for studying user watching behavior in online Internet video-on-demand (VoD) systems. It is important for system planning, user engagement understanding, and system quality evaluation. However, due to the limited access of user data in large-scale streaming systems, a systematic measurement, analysis, and modeling of video watching time is still missing. In this paper, we measure PPLive, one of the most popular commercial Internet VoD systems in China, over a three week period. We collect accurate user watching data of more than 100 million streaming sessions of more than 100 thousand distinct videos. Based on the measurement data, we characterize the distribution of watching time of different types of videos and reveal a number of interesting characteristics regarding the relation between video watching time and various video-related features (including video type, duration, and popularity). We further build a suite of mathematical models for characterizing these relationships. Extensive performance evaluation shows the high accuracy of these models as compared with commonly used data-mining based models. Our measurement and modeling results bring forth important insights for simulation, design, deployment, and evaluation of Internet VoD systems.

In this paper, we measure and study the peer startup process in PPLive, a popular commercial P2P streaming system, and focus on a fundamental issue in this aspect: how a peer initializes its buffer when it joins a channel, i.e., initial offset placement of peers' buffers in the startup stage. We build a general model of peer startup process in chunk-based P2P streaming systems and present an initial offset placement scheme we inferred from the measurement results, i.e., proportional placement (PP) scheme. With FP scheme, the initial buffer offset is set to the offset of the reference neighbor peer plus an advance proportional to the reference neighbor peer's offset lag or buffer width. We evaluate the performance of PP scheme and find it is stable when the placement is based on offset lag, but will be unstable when it is based on buffer width if the chunk fetching strategy and neighbor peer selection mechanism are not properly designed. We finally report our detailed measurement results of the peer startup process and initial offset placement algorithms used in PPLive. Our models and measurement results could be useful for guiding the analysis and design of buffering protocols for a real P2P live streaming system.

A fundamental problem that a peer-to-peer (P2P) live streaming system faces is how to support flash crowds effectively. A flash crowd occurs when a burst of join requests arrive at a system. When a flash crowd occurs, the sudden arrival of numerous peers may starve the upload capacity of a P2P system, and degrade the quality of service. By theoretical analysis and simulations, we find that a system has limited capacity to handle a flash crowd: It can recover to a new stable state when the size of flash crowd is small or moderate, but collapse when the flash crowd is excessively large. The capacity of a system is independent of initial state of the system while relevant to stable peers' departure rate, which suggests this capacity is an essential property of a P2P live streaming system. In addition, we prove that a P2P live streaming system with admission control has excellent capacity to handle flash crowds: It can recover from flash crowds of excessively large size and a startup peer's waiting time scales logarithmically with the size of flash crowds. Our theoretical model and simulation results provide a promising framework to understand the capacity of a P2P live streaming system for handling flash crowds.

In a peer-to-peer (P2P) live streaming system, peer requests collide when multiple peers request data pieces from the same peer (or media server). When collisions occur, some of the peers' requests fail and retries have to be taken, which delays peer's receipts of pieces. This paper shows that request collisions occur frequently at both media servers and peers, and have big impact on system performance. It then proposes two algorithms to address this issue. The first is a novel admission control algorithm at the media server. The second is a peer selection algorithm in which peer requests pieces from neighbors with low collision probability. Simulation results show that the proposed algorithms improve system performance significantly.

The correct evaluation of P2P streaming system models needs the validation in real world system. However, there is lack of systematic and integrated measurement method for real world P2P streaming system. In this paper, we propose a P2P streaming network measurement method based on a peer's buffer occupancy probability. Our method is based on the fixed duration buffer property of commercial P2P streaming systems. We prove the measured buffer occupancy probability reflects the chunk propagation process in the P2P network. We then propose a P2P streaming chunk propagation model and verify it in the commercial P2P streaming network using our measurement method. Our measurement method is useful for measuring and analyzing miscellaneous P2P streaming systems. And our model and parameter estimation are useful for existing P2P simulators to choose correct parameters and are meaningful for researchers to understand the real meaning behind the parameters.