Fine-Grained Sharing of Encrypted Sensor Data
Threads of data theft in public-cloud storage as well as privacy concerns are becoming more and more serious. Common wisdom has been advocating for the use of cryptographic solutions in protecting data secrecy. However, this brings forth a technical challenge on fine-grained sharing. Existing solutions such as Attribute-Based Encryption (ABE) or Key Aggregation Cryptosystem (KAC) allows key aggregation so that only one single key needs to be sent. However, they suffer from performance limitations. In this paper, we made the observation that for a large class of queries, there is a fast linear time reconstruction method, significantly reducing the reconstruction time down to the asymptotic optimal. We classify queries into three types; i.e. ranged queries, sub-sampling queries and general queries and propose approaches to achieve optimal reconstruction time for each of them. We perform extensive experiments using real data set to confirm the efficiency and practicality of our proposed approaches. Experimental studies show that our approaches are indeed indeed efficient, which outperform original KAC by upto 90X. Our technique is also practical, which is proven by its capability to reconstruct more than hundred keys per second with constrained computational power.
Micro Location Detection
Extracting useful information from tweets has been a major research topic in the areas of information extraction and natural language processing. Core tasks such as Named Entity Recognition to detect locations are well established in domains handling formal texts, but tweets are short, informal, and often poorly worded. Existing works applying NER techniques on tweets are limited to macro-locations such as countries, cities, regions, or very popular points of interest. We propose SinNER, our approach for the detection of micro-locations in tweets. For SinNER, we created a customized dictionary of location names and index them using a tailored prefix trie to lookup location candidates, resulting in a good recall. We then significantly improve the precision by training a classifier using supervised learning techniques to act as subsequent filter to distinguish between true and false locations among the set of candidates. In a comprehensive evaluation, we investigate the accuracy and performance of the different system components and show that, in terms of accuracy, SinNER clearly outperforms existing solutions in the task of detecting micro-locations.
Human Action Recognition
Intelligent video surveillance system is built to automatically detect events of interest, especially on object tracking and behaviour understanding. In this work, we focus on the task of human action recognition under surveillance environment, specifically in a multi-camera monitoring scene. Despite many approaches having achieved success in recognizing human action from video sequences, they are designed for single view and generally are not robust against viewpoint invariant. Human action recognition across different views remains challenging due to the large variations from one view to another. We present a framework to solve the problem of transferring action models learned in one view (source view) to another view (target view). Furthermore, we extend our framework to transfer action models from multiple views to one view when there are multiple source views available. Experiments on the IXMAS human action dataset, which contains videos captured with five viewpoints, show the efficacy of our framework.
Predicting Where People Look in Multiple CCTV Cameras
A significant body of literature on saliency modeling predicts where humans look at in a single image or video. Besides the scientific goal of understanding how information is fused from multiple visual sources to identify regions of interest in a holistic manner, there are tremendous engineering applications of multi-camera saliency due to the widespread installations of cameras. This paper proposes a principled framework to smoothly integrate visual information from multiple views to a global scene map, and to employ an object-aware saliency algorithm to identify the most important regions by fusing visual information. The proposed method has the following key distinguishing features compared with its counterparts:
- The proposed saliency detection is global (salient regions from one local view may not be important in a global context).
- It does not require special ways for camera deployment or overlapping fields of view.
- The key saliency algorithm is effective in highlighting interesting object regions though not a single detector is used.
Experiments on several datasets confirm the effectiveness of the proposed principled framework.
Active Camera Sensing
Active Multi-Camera Surveillance
Pan-tilt-zoom capabilities of active cameras can be exploited to provide high-quality surveillance. To achieve effective, real-time surveillance, an efficient collaborative mechanism is needed to control and coordinate these cameras’ actions. Applications like biometric task, forensic analysis, situational awareness, surveillance video mining, and many more requires to capture high resolution image of targets in surveillance.
Multimedia Data Processing
A Self-calibrating Participatory Indoor Localization System
PiLoc is an indoor localization system that utilizes opportunistically sensed data contributed by users. It does not require manual calibration, prior knowledge and infrastructure support. The key novelty of PiLoc is that it merges walking segments annotated with displacement and signal strength information from users to derive a map of walking paths annotated with radio signal strengths. As PiLoc enables minimum user effort for calibration and maintenance, it has potential for large scale deployment.
Conversations between people happen all the time in their daily lives. Conversations can takes place in many forms: face to face, emails, letters, video calls, text messages, through telephonic devices and many more. These conversations capture huge amount of information on human interactions.
Are these conversations analyzed and put to good use?
SocialWeaver is a smartphone-based conversation sensing system which can perform conversation clustering and build conversation networks among the users. It’s a user friendly application which can be downloaded onto smart devices. SocialWeaver is able to:
- Exploits conversation clustering to detect conversational interactions
- Differentiate different conversations groups • Users friendly (No training required)
- Ensure User Privacy
Measurement and Diagnosis of Internet Service Provider (ISP) Cellular Data Networks
ISP check is an crowd-sourcing based Android application that will provide consumers with an accurate view of Quality of Service that they are experiencing for their 3G/LTE networks. Various metrics are measured such as TCP/UDP download and upload speed, Round Trip Time (RTT), and DNS lookup time. Trace files are stored at servers for further analysis. Using the large amount of data collected, we can investigate the behaviour and settings of local cellular data networks, which are beneficial to design better transport layer protocols for mobile networks. Furthermore, our goal is to build a performance heatmap which monitors how networks perform everywhere in the island and help ISPs improve their networks.
The App is available to download in the Android Play Store.
Data Indexing and Analytics
GENIE and LAMP
The GENIE&LAMP project aims to provide a systematic investigation into the use of semi-lazy learning for predictive analytics. GENIE (GENeric IndEx) is an unified platform to support storage and retrieval of Big Data with various types of structure. LAMP (semi-LAzy Mining Paradigm) is a new data mining paradigm for predictive analytics, which essentially follows the lazy learning paradigm until the last step where more expensive eager learning methods are applied on a set of KNNs. Specifically, GENIE will efficiently support generic k-Nearest Neighbours (k-NN) search which will be utilized by LAMP to empower users with advanced predictive capabilities.
Hybrid Cloud VS
Processing of Mixed-sensitivity Sensing Data in Hybrid Clouds
Video surveillance systems generate huge volumes of video streams which often contain both sensitive and non-sensitive data. How to process this large-scale mixed-sensitivity video data in both secure and efficient manner?