The advent of in-band network telemetry (INT) facilitates the dynamic and fine-grained monitoring of network conditions. Nevertheless, the current INT specification incurs a substantial measurement overhead, diminishing bandwidth utilization, and potential data fragmentation. Moreover, the rapid accumulation of telemetry reports may give rise to data lakes in the collector, resulting in computational burden and telemetry degradation. To address these two problems, we propose an integrated solution for high-efficiency in-band network telemetry, including an improved INT scheme, SF-INT, to reduce measurement overhead and the usage of SmartNIC to accelerate report processing. SF-INT employs network node registers to store the telemetry information collected at previous nodes by previous packets. When a packet arrives at one node, the node performs store-and-forward actions to determine the telemetry information stored in the current register and the telemetry information carried by the packet to the next hop. The packet carries only one node’s telemetry information during a telemetry process, thus keeping a constant packet size. In contrast, the standard INT increases the packet size due to the hop-by-hop insertion of metadata. SmartNIC offloads report processing from the CPU and obtains comprehensive telemetry data by analyzing a set of telemetry reports. The testbed experiment results have demonstrated that our solution reduces measurement overhead and efficiently processes reports at nearly the line rate.

Traditional network measurements mainly focus on end-to-end performance. In-band Network Telemetry (INT) makes a change by embedding telemetry instructions in the packet header and obtaining the status information of each network device. This information forms telemetry metadata forwarded with the packet hop by hop, such that we can observe network conditions more comprehensively. However, INT cannot control the route path or focus on the measurement of critical nodes, while the measurement of all nodes costs a lot. This paper proposes INToSR, a novel In-band Network Telemetry that fully leverages the flexible path programming capabilities of Segment Routing over IPv6 (SRv6). It first defines new SRv6 Endpoint Behaviors and puts INT instructions in the Arguments field. Thus, INToSR enables path control by setting an INT path using the SegmentList of SRv6 and monitors the metrics of critical nodes like queue length, link workload, and packet latency, achieving fine-grained network measurement. INToSR is also lightweight because it reduces the embedded telemetry metadata, the header overhead, and the processing delay. An experimental testbed has been built, including Tofino chips, which implement INToSR in the data plane, as well as the SDN controller and the INT report processing platform in the control plane. The experiment results have verified the feasibility of INToSR, which achieves the nanosecond-level network measurement and outperforms two existing methods. Finally, the accuracy of INToSR is evaluated in locating network congestion, which provides references for optimizing network performance.

Website Fingerprinting (WF) attacks enable passive adversaries to identify the website a user visits over encrypted or anonymized network connections. WF attacks based on deep learning have achieved high accuracy in identifying websites based on abundant training traffic traces per website. However, collecting large-scale and fresh traces is quite cost-consuming and unrealistic. Morevoer, these deep-learning-based WF attacks lack flexibility because they require a long bootstrap time for retraining when facing new traffic traces with different distributions or newly added monitored websites. This paper proposes a high-accuracy WF attack named Contrastive Fingerprinting (CF), which leverages contrastive learning and data augmentation over a few training traces. The results of extensive experiments on challenging datasets over few-shot traces demonstrate the high accuracy of the CF attack and its robustness against WF defenses. For example, when each monitored website only has 20 training traces, CF identifies monitored websites with a high accuracy of 90.4% in the closed-world scenario and distinguishes monitored websites with a high True Positive Rate of 91.2% in the open-world scenario. The experimental results also show that CF outperforms two existing WF attacks with few-shot traces under different network conditions in real-world applications.

Many web applications adopt Transfer Control Protocol (TCP) as the underlying protocol, where congestion control (CC) plays a vital role in reliable transmission. However, some TCP mechanisms cannot cope with the requirements of new applications and ever-increasing network traffic. Therefore, people have proposed Quick UDP Internet Connection (QUIC), an excellent potential alternative based on UDP, which introduces new features to improve transmission performance and is compatible with existing CC algorithms. This paper has conducted many experiments in the testbed and actual environments to measure and compare QUIC and TCP regarding communication quality, compatibility fairness, and user experience, while considering the impacts of three typical CC algorithms: NewReno, Cubic, and BBR. QUIC outperforms TCP in most experiments for web browsing and online video, but its performance is susceptible to CC algorithms and network conditions. For example, with the Cubic algorithm, QUIC enabling the 0-RTT feature can decrease the webpages loading time by 37.11% compared with TCP. Using the BBR algorithm, both QUIC and TCP achieve high throughput, slight fluctuation, and few delayed events when playing online videos. TCP with BBR provides better fairness, while QUIC with BBR is more robust in a network with high latency or packet loss.

Website Fingerprinting (WF) enables a passive adversary to identify the website a user is visiting, even when the web access adopts security or privacy technologies. WF attacks based on deep learning are highly effective when feeding sufficient training traces, for example, hundreds of traffic traces of accessing each website. However, collecting extensive traffic consumes much time and resources, degenerating WF attacks' timeliness and invisibility. Nevertheless, decreasing training traces dramatically drops the WF accuracy. This paper proposes Tripod, a novel data augmentation method to enhance WF attacks, making them effective with a small training set. It applies three packet manipulations (Injecting, Removing, and Losing) on one collected traffic trace to generate several augmented traces. WF attacks then use the website classifier trained by the augmented set of all traces. In the closed-world scenario, the Var-CNN attack with 20 training traces per website only correctly identifies 56.1% of websites, while Tripod significantly increases this accuracy to 95.9%. Furthermore, Tripod increases the true positive rate of Var-CNN from 26.9% to 91.4% in the more realistic open-world scenario.

Online learned tracking is widely used to handle the appearance changes of object because of its adaptive ability. Learning to rank technique has attracted much attention recently in visual tracking. But the tracking method with online learning to rank suffers from the error accumulation problem during the self-training process. To solve this problem, we propose an online learning to rank algorithm in the co-training framework for robust visual tracking. A co-training algorithm combined with ranking SVM collects features and unlabeled data for training. Two ranking SVMs are built with different types of features accordingly and dynamically fused into a semi-supervised learning process. This semi-supervised learning approach is updated online to resist the occlusion and adapt to the changes of object's appearance. Many experiments on challenging sequences have shown that the proposed algorithm is more effective than the state-of-the-art methods.

The sparse representation has been widely used in many areas including visual tracking. The part-based representation performs outstandingly by using non-holistic templates to against occlusion. This paper combined them and proposed a robust object tracking method using part-based sparsity model for tracking an object in a video sequence. In the proposed model, one object is represented by image patches. The candidates of these patches are sparsely represented in the space which is spanned by the patch templates and trivial templates. The part-based method takes the spatial information of each patch into consideration, where the vote maps of multiple patches are used. Furthermore, the update scheme keeps the representative templates of each part dynamically. Therefore, trackers can effectively deal with the changes of appearances and heavy occlusion. On various public benchmark videos, the abundant results of experiments demonstrate that the proposed tracking method outperforms many existing state-of-the-arts algorithms.

Real time visual tracking is a challenge problem in computer vision. In this paper, we propose a real-time tracking method based on compressive sensing Markov Chain Monte Carlo (MCMC) sampling. To extract the features of objects, non-adaptive random projections are employed in the object appearance model which adopts a very sparse random measurement matrix using compress sensing. These projection preserve the structure of objects in the image feature space. A Bayesian classifier is learnt from the object appearance model and the scores of this classifier are integrated into Markov Chain Monte Carlo acceptance mechanism. Furthermore, a two-stage tracking scheme is used to alleviate the drift problem. The experimental results demonstrate that the proposed method is real time and outperforms some start-of-the-art algorithms on public benchmark sequences in terms of accuracy and robustness.

The well-built dataset is a pre-requisite for object categorization. The processes of collecting and labeling the images are laborious and monotonous. In order to label images efficiently, we propose an incremental learning model to label images automatically with a bounding box for each visual object category. Our approach combines image classification and object detection. Given a query image, our approach firstly searches the candidate regions coarsely using the beyond sliding windows scheme, and then locate the object finely using the sliding window scheme, and after that update the learning model. We use two criteria to evaluate the image labeling, the detection precision and the detection consistency with the ground truth label. Our approach can localize the object fast and sequentially update the learning model with the increasing of the unlabeled samples. The experiment results have demonstrated that our approach outperforms the BOW model in terms of the precision and consistency of detection.

This paper aims to recognize ID numbers from three types of valid identification documents in China: the first-generation ID card, the second-generation ID card and the driver license of motor vehicle. We have proposed an approach using local similarity voting to automatically recognize ID numbers. Firstly, we extract the candidate region which contains ID numbers and then locate the numbers and characters. Secondly, we recognize the numbers by an improved template matching method based on the local similarity voting. Finally, we verify the ID numbers and characters. We have applied the proposed approach to a set of 100 images which are shot by conventional digital cameras. The experimental results have demonstrated that this approach is efficient and is robust to the change of illumination and rotation. The recognition accuracy is up to 92%.

Face alignment is a significant problem in the processing of face image. Active Shape Model (ASM) is one of powerful solution to this problem. But ASM is sensitive to the localization of initial points. If the initial shape is not located in the appropriate location, the ASM iteration may not converge. In this paper, in order to improve the ASM convergence, we proposed an approach for face alignment based on the face component localization. We firstly use the component localization for locating initial points, and then match the landmark points with the face appearance features. Such that the ASM can converge fast with the less errors of iterations since that the component localization provides good initial points in the preprocessing. We evaluated our approach on the frontal views of upright faces of IMM dataset. The experimental results have shown that our approach outperforms the original ASM in terms of efficiency and accuracy.

The problem of object tracking in dense clutter is a challenge in computer vision. This paper proposes a method for tracking object robustly by combining the online selection of discriminative color features and the offline selection of discriminative Haar features. Furthermore, the cascade particle filter which has four stages of importance sampling is used to fuse two kinds of features efficiently. When the illumination changes dramatically, the Haar features selected offline play a major role. When the object is occluded, or its rotation angle is very large, the color features selected online play a major role. The experimental results show that the proposed method performs well under the conditions of illumination change, occlusion, object scale change and abrupt motion of object or camera.