We describe our current work in developing novel mechanisms for managing security and privacy in pervasive computing environments. More specifically, we have developed and evaluated three different applications, including a contextual instant messenger, a people finder application, and a phone-based application for access control. We also draw out some themes we have learned thus far for user-controllable security and privacy. 1.

As statistical machine learning algorithms and techniques continue to mature, many researchers and developers see statistical machine learning not only as a topic of expert study, but also as a tool for software development. Extensive prior work has studied software development, but little prior work has studied software developers applying statistical machine learning. This paper presents interviews of eleven researchers experienced in applying statistical machine learning algorithms and techniques to human-computer interaction problems, as well as a study of ten participants working during a five-hour study to apply statistical machine learning algorithms and techniques to a realistic problem. We distill three related categories of difficulties that arise in applying statistical machine learning as a tool for software development: (1) difficulty pursuing statistical machine learning as an iterative and exploratory process, (2) difficulty understanding relationships between data and the behavior of statistical machine learning algorithms, and (3) difficulty evaluating the performance of statistical machine learning algorithms and techniques in the context of applications. This paper provides important new insight into these difficulties and the need for development tools that better support the application of statistical machine learning.

In order to develop intelligent systems that attain the trust of their users, it is important to understand how users perceive such systems and develop those perceptions over time. We present an investigation into how users come to understand an intelligent system as they use it in their daily work. During a six-week field study, we interviewed eight office workers regarding the operation of a system that predicted their managers ’ interruptibility, comparing their mental models to the actual system model. Our results show that by the end of the study, participants were able to discount some of their initial misconceptions about what information the system used for reasoning about interruptibility. However, the overarching structures of their mental models stayed relatively stable over the course of the study. Lastly, we found that participants were able to give lay descriptions attributing simple machine learning concepts to the system despite their lack of technical knowledge. Our findings suggest an appropriate level of feedback for user interfaces of intelligent systems, provide a baseline level of complexity for user understanding, and highlight the challenges of making users aware of sensed inputs for such systems. Author Keywords Intelligent systems, context-aware, mental models,

We present a novel system for notification management and report results from two studies testing its performance and impact. The system uses statistical models to realize defer-to-breakpoint policies for managing notifications. The first study tested how well the models detect three types of breakpoints within novel task sequences. Results show that the models detect breakpoints reasonably well, but struggle to differentiate their type. Our second study explored effects of managing notifications with our system on users and their tasks. Results showed that scheduling notifications at breakpoints reduces frustration and reaction time relative to delivering them immediately. We also found that the relevance of notification content determines the type of breakpoint at which it should be delivered. The core concept of scheduling notifications at breakpoints fits well with how users prefer notifications to be managed. This indicates that users would likely adopt the use of notification management systems in practice.

Abstract. We describe the design of privacy controls and feedback mechanisms for contextual IM, an instant messaging service for disclosing contextual information. We tested our designs on IMBuddy, a contextual IM service we developed that discloses contextual information, including interruptibility, location, and the current window in focus (a proxy for the current task). We deployed our initial design of IMBuddy’s privacy mechanisms for two weeks with ten IM users. We then evaluated a redesigned version for four weeks with fifteen users. Our evaluation indicated that users found our group-level rulebased privacy control intuitive and easy to use. Furthermore, the set of feedback mechanisms provided users with a good awareness of what was disclosed.

Abstract. This paper addresses the problem of capturing rich, longterm personal activity logs of users ’ interactions with their workstations, for the purpose of deriving predictive, personal user models. Our architecture addresses a number of practical problems with activity capture, including incorporating heterogeneous information from different applications, measuring phenomena with different rates of change, efficiently scheduling knowledge sources, incrementally evolving knowledge representations, and incorporating prior knowledge to combine low-level observations into interpretations better suited for user modeling tasks. We demonstrate that the computational and memory demands of general activity capture are well within reasonable limits even on today’s hardware and software platforms. 1

As machine learning (ML) systems emerge in end-user applications, learning algorithms and classifiers will need to be robust to an increasingly unpredictable operating environment. In many cases, the parameters governing a learning system cannot be optimized for every user scenario, nor can users typically manipulate parameters defined in the space and terminology of ML. Conventional approaches to user-oriented ML systems have typically hidden this complexity from users by automating parameter adjustment. We propose a new paradigm, in which model and algorithm parameters are exposed directly to end-users with intuitive labels, suitable for applications where parameters cannot be automatically optimized or where there is additional motivation – such as creative flexibility – to expose, rather than fix or automatically adapt, learning parameters. In our CHI 2008 paper, we introduced and evaluated MySong, a system that uses a Hidden Markov Model to generate chords to accompany a vocal melody. The present paper formally describes the learning underlying MySong and discusses the mechanisms by which MySong‟s learning parameters are exposed to users, as a case study in making ML systems user-configurable. We discuss the generalizability of this approach, and propose that intuitively exposing ML parameters is a key challenge for the ML and human-computer-interaction communities. 1. Introduction and Related

A notification represents the proactive delivery of information to a user and reduces the need to visually scan or repeatedly check an external information source. At the same time, notifications often interrupt user tasks at inopportune moments, decreasing productivity and increasing frustration. Controlled studies have shown that linking notification delivery to the perceptual structure of a user’s tasks can reduce these interruption costs. However, in these studies, the scheduling was always performed manually, and it was not clear whether it would be possible for a system to mimic similar techniques. This article contributes the design and implementation of a novel system called Oasis that aligns notification scheduling with the perceptual structure of user tasks. We describe the architecture of the system, how it detects task structure on the fly without explicit knowledge of the task itself, and how it layers flexible notification scheduling policies on top of this detection mechanism. The system also includes an offline tool for creating customized statistical models for detecting task structure. The value of our system is that it intelligently schedules notifications, enabling the reductions in interruption costs shown within prior controlled studies to now be realized by users in everyday desktop computing tasks. It also provides a test bed for experimenting with how notification management policies and other system functionalities can be linked to task

Abstract. Establishing secure communication channels between devices that share no a priori context, also known as the device-pairing problem, is the first step towards the realization of mobile, interoperable applications interacting across several devices. One approach that promises to be both secure and usable relies on so-called auxiliary or out-of-band channels for authentication. Although many such solutions have been independently suggested, open, easily usable implementations are unfortunately missing so far. In this demonstration, we present OpenUAT, an open source toolkit for establishing secure ad-hoc connections. Open-UAT implements many of the auxiliary channels proposed in the past few years on top of a unified, common cryptographic protocol for key exchange and runs on a variety of mobile phones and desktop/laptop computers. By giving users the chance to directly compare different device pairing alternatives in a real-life security prototype, OpenUAT fosters usability research and shortens the gap between research prototypes and real-world applications. 1

My research focuses on human-centered approaches to developing, deploying, and evaluating sensor-based interfaces in everyday life. As a human-computer interaction researcher, I am interested in user interface software and technology, in ubiquitous computing, and in the human-centered development of statistical models. Sensor-Based Statistical Models of Human Interruptibility My dissertation work is in the domain of human interruptibility. While a colleague glancing into an office can generally make a rapid determination of whether it is currently appropriate to interrupt, current systems are largely blind to the situations surrounding their usage. Current systems therefore tend to interrupt at inappropriate times or unduly demand attention. In a series of studies, I examine the feasibility and robustness of sensor-based statistical models of human interruptibility. I show that models learned from practical sensors can identify “highly non-interruptible ” situations better than human observers. Informed by these studies, I create a tool to enable non-expert development of applications that sense and model human situations. Sensor development can be costly and time-consuming, and it was unclear what sensors