Edge Analytics in the Internet of Things

Smart Spaces

High-data-rate sensors are becoming ubiquitous in the Internet of Things. GigaSight is an Internet-scale repository of crowd-sourced video content that enforces privacy preferences and access controls. The architecture is a federated system of VM-based cloudlets that perform video analytics at the edge of the Internet.

any of the “things” in the might be seen in the video of someone record- Internet of Things (IoT) are ing the event.7 Surveillance videos were crucial video cameras, which are for discovering the Boston Marathon bombers proliferating at an astonish- in 2013. In general, an image captured for one ing rate. In 2013, it was esti- reason can be valuable for some totally unre- Mmated that there was one surveillance cameras lated reason. Stieg Larsson’s fictional work, The for every 11 people in the UK.1 Video cameras Girl with the Dragon Tattoo (Alfred A. Knopf, are now common in police 2010), embodies exactly this theme: a clue to patrol cars2,3 and almost solving a crime is embedded in the backgrounds Mahadev Satyanarayanan universal in Russian passen- of old crowd-sourced photographs. Carnegie Mellon University ger cars.4 The emergence of This richness of content and the possibil- commercial products, such ity of unanticipated value distinguishes video Pieter Simoens as Google Glass and GoPro, from simpler sensor data that has historically Ghent University point to a future in which been the focus of the sensor network research Yu Xiao body-worn video cameras community. The sidebar presents many hypo- Aalto University will be commonplace. Many thetical use cases for crowd-sourced video. An police forces in the US are now Internet-scale searchable repository for crowd- Padmanabhan Pillai considering the use of body- sourced video content, with strong enforce- Intel Labs worn cameras.5 The report of ment of privacy preferences and access controls, Zhuo Chen, Kiryong Ha, the 2013 NSF Workshop on would be a valuable global resource. Here, we Wenlu Hu, and Brandon Amos Future Directions in Wireless examine the technical challenges involved in Carnegie Mellon University Networking predicts that “it creating such a repository. In particular, we will soon be possible to find a propose GigaSight, a hybrid cloud architecture camera on every human body, that is effectively a content delivery network in every room, on every street, (CDN) in reverse. and in every vehicle.”6 What will we do with all this video? To- GigaSight: A Reverse CDN day, most video is stored close to the point of A key challenge for the cloud is the high cumu- capture, on local storage. Its contents are not lative data rate of incoming videos from many easily searched over the Internet, even though cameras. Without careful design, this could there are many situations in which timely re- easily overwhelm metropolitan area networks mote access can be valuable. For example, at a (MANs) and ingress Internet paths into a cen- large public event such as a parade, a lost child tralized cloud infrastructure, such as Google’s

ere we present some hypothetical use cases for crowd- When a dog owner reports that his dog is missing, a search of H sourced video. crowd-sourced videos captured in the last few hours may help locate the dog before it strays too far. Marketing and Advertising Crowd-sourced videos can provide observational data for ques- Public Safety tions that are difficult to answer today. For example, which Where are the most dangerous intersections, with an accident billboards attract the most user attention? How successful is a waiting to happen? Perhaps an accident hasn’t happened yet, new store window display in attracting interest? Which clothing but it could just be a matter of time. Video analytics of intersec- colors and patterns attract the most interest? Are there regional tions could lead to the timely installation of traffic lights. preferences? Many other public safety improvements are also possible: uneven sidewalks that are causing people to trip and fall; timely Theme Parks detection of burned-out street lights that need to be replaced, Visitors to places like Disneyworld can capture and share their new potholes that need to be filled, and icy road surfaces and experiences, including rides, throughout the entire day. With sidewalks in need of immediate attention. video, audio, and accelerometer capture, the recreation of rides can be quite realistic. An album of a family’s visit can be shared Fraud Detection via social networks such as Facebook or Google+. A driver reports that his car was hit while it was parked at a restaurant. However, his insurance claims adjuster finds a crowd- Locating People, Pets, and Things sourced video in which the car is intact when leaving the A missing child was last seen walking home from school. A search restaurant. of crowd-sourced videos from the area shows that the child was Other law and order opportunities abound. For example, near a particular spot an hour ago. The parent remembers that when a citizen reports a stolen car, his description could be used the child has a friend close to that location. She calls the friend’s to search recent crowd-sourced video for sightings of that car to home and locates the child. help locate it.

large datacenters or Amazon’s Elastic Compute Cloud (EC2) sites. In 2013, roughly one hour’s worth of video Global catalog of video segments was uploaded to YouTube each sec- Video ID Cloudlet IP address Owner Start time End time Index terms... ond. That corresponds to 3,600 concurrent uploads. Scaling well beyond this to millions of concurrent uploads Cloud from a dense urban area will be difficult. Today’s high-end MANs only have a bandwidth of 100 Gbps. Each such link can support 1080p streams Internet from only 12,000 users at YouTube’s recommended upload rate of 8.5 Mbps. A million concurrent uploads would re- Cloudlet 1 Cloudlet 2 quire 8.5 Tbytes per second. VM1 VM1 To solve this problem, we propose VM VM2 Denatured 2 Denatured GigaSight, a hybrid cloud architecture videos videos VM VM3 that uses a decentralized cloud com- 3 puting infrastructure in the form of virtual machine (VM)-based cloud- Figure 1. The GigaSight architecture. It uses a decentralized cloud computing lets (see Figure 1).8 A cloudlet is a new infrastructure in the form of VM-based cloudlets. architectural element that arises from

April–june 2015 PERVASIVE computing 25 Smart Spaces

the convergence of mobile computing are unloaded. The potential bandwidth recognition in individual frames. Ac- and cloud computing. It represents the bottleneck is at the access and aggrega- tivity recognition in individual se- middle of a three-tier hierarchy: mobile tion networks and not in the core net- quences is also possible. However, device, cloudlet, and cloud. A cloud- work with its high-speed links. preserving privacy involves more than let can be viewed as a “data center in Preprocessing videos on cloudlets blurring (or completely removing) a box” that “brings the cloud closer.” also offers the potential of using con- frames with specific faces, objects, or Although cloudlets were originally tent-based storage optimization al- scenes in the personal video. By using created to address end-to-end latency gorithms to retain only one of many other objects in the scene, or by com- in interactive applications, the use of similar videos from colocated cameras. paring videos from other users taken cloudlets in GigaSight is based solely Thus, cloudlets are the true home of at the same place or time but with dif- on bandwidth considerations. denatured videos. In a small number ferent privacy settings, we might still In the GigaSight architecture (de- of cases, based on popularity or other deduce which object was blurred and scribed in detail elsewhere9), video metrics of importance, some videos can thus of value to the person who cap- from a mobile device only travels as be copied to the cloud for archiving or tured the video. The user’s denaturing far as its currently associated cloudlet. replicated in the cloud or other cloud- policy must also be applied to videos Computer vision analytics are run on lets for scalable access. But most videos that were captured by others at ap- the cloudlet in near real time, and only will reside only at a single cloudlet for proximately the same time and place. the results (recognized objects, rec- a finite period of time (typically on the Simply sending the denaturing rules ognized faces, and so on), along with order or hours, days, or weeks). In a to the personal VMs of other parties metadata (such as the owner, capture commercial deployment of GigaSight, is undesirable; this would expose at a location, and timestamp) are sent to how long videos remain accessible will metalevel the sensitive content. the cloud. The tags and metadata in depend on the storage retention and One possible solution, proposed by the cloud can guide deeper and more billing policies. Jianping Fan and colleagues,10 is to customized searches of the content of a Note that Figure 1 is agnostic regard- send weak object classifiers to a cen- video segment during its (finite) reten- ing the exact positioning of the cloud- tral site where they are combined with tion period on a cloudlet. lets in the network. One option is to a global concept model. This model An important type of “analytics” place numerous small cloudlets at the could then be returned to the personal supported on cloudlets is automated network edge. An alternative is to place VMs. Of course, this approach requires modification of video streams to pre- fewer but larger cloudlets deeper in the videos to be temporarily saved in the serve privacy. For example, this might network—for example, at the metro- personal VM until the central site has involve editing out frames or blurring politan scale. Our analysis suggests received any video uploaded at the same individual objects within frames. What that small cloudlets close to the edge is time and place. Any video that is up- needs to be removed or altered is highly the better alternative.9 loaded later could then simply be dis- specific to the owner of a video stream, carded, to avoid keeping videos in the but no user has time to go through and Denaturing personal VM for too long. manually edit video captured on a con- Denaturing must strike a balance be- In its full generality, denaturing tinuous basis. This automated, owner- tween privacy and value. At one ex- might involve not only content modifi- specific lowering of fidelity of a video treme is a blank video: perfect privacy cation but also metadata modification. stream to preserve privacy is called de- but zero value. At the other extreme is For example, the accuracy of location naturing (discussed further in the next the original video at its capture resolu- metadata associated with a sequence section). tion and frame rate. This has the high- of video frames might be lowered to It is important to note in Figure 1 that est value for potential customers, but meet the needs of k-anonymity in loca- cloudlets are not just temporary staging it also incurs the highest exposure of tion privacy.11 Whether the contents of points for denatured video data in route privacy. Where to strike the balance the video sequence will also have to be to the cloud. With a large enough num- is a difficult question that is best an- blurred depends on the video’s visual ber of cameras and continuous video swered individually, by each user. distinctiveness—a scene with the Eiffel capture, the constant influx of data at This decision will most probably be Tower in the background is obviously the edges will be a permanent stress on context-sensitive. locatable even without explicit location the ingress paths to the cloud. Just buff- Denaturing is a complex process metadata. ering data at cloudlets for later trans- that requires careful analysis of the In the future, guidance for denatur- mission to the cloud won’t do. Because captured frames. State-of-the-art com- ing might also be conveyed through so- video will be streaming 24/7, there will puter-vision algorithms enable face cial norms that deprecate video capture never be a “later” when ingress paths detection, face recognition, and object of certain types of scenes or in certain

26 PERVASIVE computing www.computer.org/pervasive locations. A system of tagging locations or objects with visual markers (such as QR codes) could indicate that video Cloud capture is unwelcome. We can imagine video capture devices that automati- cally refrain from recording when they Video metadata recognize an appropriate QR code in the scene. In addition, denaturing al- Indexer VM Tagging of gorithms on a cloudlet might also strip Cloudlet denatured video Thumbnail videos out scenes that contain such a code. In Encrypted segments storage the long run, we can envision the emer- Denatured segments gence of an ethics of video capture in public. Many broader societal issues, Personal VM1 such as the ability to subpoena cap- Cloudlet tured but encrypted video, add further Video complexity. Clearly, denaturing is a Stream 1 Encoder

deep concept that will need time, effort, Decoder Content filter and deployment experience to fully un- Sampling Early discard Blanking Blurring derstand. GigaSight opens the door to Personal VM2 exploring these issues and to evolving a Personal VM4 Personal VM6 decoder encoder content filter sampling early discard blanking blurring encoder societally acceptable balance between decoder content filter sampling early discard blanking blurring encoder decoder content filter privacy and utility. Personal VM3 sampling early discard blanking blurring In the GigaSight prototype, a per- Personal VM5 encoder decoder content filter early discard Other video stream s sampling blanking blurring • • • encoder sonal VM on the cloudlet denatures decoder content filter sampling early discard blanking blurring each video stream in accordance with its owner’s expressed preferences. This VM is the only component, apart from the mobile device itself, that accesses Figure 2. Cloudlet processing in the GigaSight prototype. Denaturing is the original (nondenatured) video. Fig- implemented as a multistep pipeline. ure 2 illustrates the processing within a personal VM. Denaturing is implemented as a multistep pipeline. that segment. This newly denatured by a separate VM on a cloudlet. To In the first step, a subset of the video video segment can then be cached for handle searches that are time-sensi- frames is selected for actual denatur- future reuse. tive (such as locating a lost child) or to ing. Our initial experiments showed After sampling video frames, meta- search for content that is not indexed, that denaturing is too compute-in- data-based filters with low computa- custom search code encapsulated in a tensive to perform at the native video tional complexity are applied. This VM can directly examine denatured frame rate.9 Consequently, the denatur- early-discard step is a binary process: videos. For each tag produced by the in- ing process results in two output files: based on the time, location, or other dexer, an entry is created in a dedicated a low frame-rate “thumbnail video” metadata, the frame is either completely tag table of the cloudlet database. Each file that provides a representative over- blanked or passed through unmodified. entry contains the tag, the ID of the view of video content for indexing and Then, we apply content-based filters video segment, and a confidence score. search operations, and an encrypted that are part of the preference specifica- For example, an entry “dog, zoo.mp4, version of the original video. Both out- tions for denaturing. For example, face 328, 95” indicates that our indexer de- puts are stored on the cloudlet, outside detection and blurring using specified tected with 95 percent confidence a dog the personal VM. The encryption of code within the personal VM might be in frame 328 of the video zoo.mp4. Af- the full-fidelity video uses a per-session performed on each frame. Figure 3 il- ter extraction, these tags are also prop- AES-128 private key that is generated lustrates the output of such a denatured agated to the catalog of video segments by the personal VM. If a search of the frame. in the cloud. thumbnail video suggests that a partic- The throughput of indexing depends ular segment of the full-fidelity video Indexing and Search on the number of objects that must be might of interest, its personal VM can The indexing of denatured video con- detected. Because this number is po- be requested to decrypt and denature tent is a background activity performed tentially very high, we propose first

April–june 2015 PERVASIVE computing 27 Smart Spaces

a school outing to the Carnegie Science Center in Pittsburgh, showing two children in a room full of yellow balls and one of the children wearing his favorite blue striped shirt.” The first step of the search would use the time and location information and the “face” tag to narrow the search. The result is a potentially large set of thumbnails from denatured videos that cover the specified location. From a multihour period of video capture by all visitors, this might only narrow the search to a few hun- dred or few thousand thumbnails. Us- ing a color filter tuned to yellow, fol- lowed by a composite color/texture filter tuned to blue and striped most of these thumbnails can be discarded. Only the few thumbnails that pass this entire bank of filters are presented to Figure 3. An example of a denatured video frame, based on the following query: the user. From this small set of thumb- “Images taken yesterday between 2pm and 4pm during a school outing to the nails, it is easy for the user to pick the Carnegie Science Center in Pittsburgh, showing two children in a room full of yellow result shown in Figure 3. balls and one of the children wearing his favorite blue striped shirt.” GigaSight only processes video that is voluntarily shared. Datasets gath- applying only classifiers for the most Viewing all of the video segments ered via crowd-sourcing often exhibit popular objects sought in the database. identified in the first step might over- a sampling bias toward popular events Classifiers of less popular objects could whelm the user. We therefore perform or news, and the “long tail” is much be applied on an ad hoc basis if needed. a second search step that filters based less covered. We believe this content As a proof of concept, the GigaSight on actual content to reduce the re- bias will be lower in GigaSight, because prototype uses a Python-based imple- turned results to a more relevant set. many of its data sources (police on- mentation of Jamie Shotton and his This step is computationally intensive body cameras, automobile dashboard colleagues’ image categorization and but can run in parallel on the cloud- cameras, and so on) involve continuous segmentation algorithm,12 with clas- lets. This step uses early discard, as de- capture of video. Conversely, pruning sifiers trained on the MSRC21 dataset scribed by Larry Huston and his col- the collection of videos at locations and mentioned in their work. This enables leagues,13 to increase the selectivity of times, where much redundant footage the detection and tagging of 21 classes a result stream. Using a plug-in inter- is available, would limit the richness of of common objects such as airplanes, face, image-processing code fragments, data collected by GigaSight. An “un- bicycles, birds, and boats. called filters, can be inserted into the interesting” detail that is eliminated in GigaSight uses a two-step hierarchi- result stream. These code fragments let the pruning could be exactly the crucial cal workflow to help a user find video user-defined classifiers examine video evidence for an important future inves- segments relevant to a specific context. segments and discard irrelevant parts tigation, such as one of the scenarios in First, the user performs a conventional of them, thus reducing the volume of the sidebar. SQL search on the cloud-wide catalog. data presented to the user. We provide The query might involve metadata such a suite of filters for common search at- Automotive Environments as time and location, as well as tags ex- tributes, such as color patches and tex- The GigaSight architecture is especially tracted by indexing. The result of this ture patches. For more complex image relevant to automobiles. For the fore- step is a list of video segments and their content, the user can train his or her seeable future, cloud connectivity from denatured thumbnails. The identity own filters offline and insert them into a moving automobile will be 3G or 4G. (the host names or IP addresses) of the the result stream. An important question is whether cloud- cloudlets on which those video seg- To illustrate this two-step workflow, lets should be placed in automobiles or ments are located can also be obtained consider a search for “any images taken at cell towers. We see value in both al- from the catalog. yesterday between 2pm and 4pm during ternatives, as shown in Figure 4. This

28 PERVASIVE computing www.computer.org/pervasive architecture can be viewed as a mapping of Figure 1 to the automotive context. Continuous capture and real-time Today’s unmodiﬁed analytics of car-mounted video cameras cloud can help to improve road safety. For example, if the computer vision analytics on your automobile’s cloudlet recog- nizes a pothole, dead animal, or fallen tree branch, it can transmit the coordi- nates of the hazard (including a brief Internet video segment) to its cell tower cloudlet. The cloudlet can share this informa- 4G base tion promptly with other cars associ- Cloudlet 4G base Cloudlet station station ated with that cell tower. With advance warning of the hazard, those cars can proactively shift lanes to avoid the haz- Zone 1Zone 2 ard. Such transient local knowledge can also be provided when an automobile first associates with a cell tower. There Cloudlet Cloudlet Cloudlet Cloudlet Cloudlet Cloudlet is a subtle trust issue implicit in this ca- pability: a hazard report is assumed to Figure 4. GigaSight for cars. Continuous capture and real-time analytics of car- be truthful. Requiring a video segment mounted video cameras can help to improve road safety when shared via cell tower to support the report can help increase cloudlets. confidence, but it is difficult to verify when and where that video segment was captured. This trust issue points frames that each user contributes per to the need for certified sensor output, unit of time (F).9 This conceptual Network Compute N bound bound where the provenance and value of the tradeoff is illustrated in Figure 5. For 14,15 sensed data is not in doubt. values of F < F , the number of users 1 E ~ — An automobile cloudlet could also supported is limited to NE users. For F (N ,F ) perform real-time analytics of high- values F > FE, the architecture is com- E E data-rate sensor streams from the en- pute bound and N < NE. gine and other sources, alerting the Using measurements from the Gi- Users per cloudlet Feasible region driver of imminent failure or to the gaSight prototype and extrapolating need for preventive maintenance. In hardware improvements over a five- Processed frames F per second per user addition, such information can also be year timeframe, the analysis compares transmitted to the cloud for integration the two alternative design strategies, Figure 5. Cloudlet sizing tradeoff. This into a database maintained by the vehi- illustrated in Figure 6. In Figure 6a, figure illustrates the tradeoff between cle manufacturer. Fine-grain analysis of many small cloudlets are deployed close the number of users (N) and the such anomaly data might reveal model- to the edge of the Internet. In Figure processed framerate per user (F). The specific defects that can be corrected in 6b, a single large cloudlet covers a city- shaded region represents the range of a timely manner. sized area. The analysis concludes that feasible choices with fixed network and edge cloudlets (Figure 6a) scale better processing capacity. Cloudlet Size and Placement than MAN cloudlets (Figure 6b) from The scalability of GigaSight depends the viewpoint of performance. In five on the specific configurations of cloud- years, the analysis estimates that one somewhat larger than that suggested by lets and their locations in the Internet. edge cloudlet would be able to support performance considerations alone. We have analyzed the tradeoff between approximately 120 users with real-time cloudlet computational capacity and denaturing and indexing at the rate of the number of cloudlets, with the goal 30 frames per second. However, be- he central theme of Giga of maximizing both the number of si- cause management costs decrease with Sight is processing edge- multaneous users per cloudlet (N) and centralization, a more holistic analysis sourced data close to the the number of denatured and indexed might suggest an optimum size that is T point of capture in space

April–june 2015 PERVASIVE computing 29 Smart Spaces

These are just a few of examples of IoT data sources that have high data rates. Many of the points made in this article apply directly to this broad CE Core network class of sensors.

IE Metropolitan Acknowledgments area network This article is based on material that originally appeared in “Scalable Crowd-Sourcing of Video from Mobile Devices.”9 This research was supported by the US National Science Foundation (NSF) under grant number IIS-1065336, by an (a) Intel Science and Technology Center grant, by DARPA Contract No. FA8650-11-C-7190, and by the US Department of Defense (DoD) under Con- tract No. FA8721-05-C-0003 for the operation of the Software Engineering Institute (SEI), a feder- Metropolitan ally funded research and development center. area network This material has been approved for public release Core and unlimited distribution (DM-0000276). Ad- network ditional support was provided by IBM, Google, Bosch, Vodafone, and the Conklin Kistler fam- I ily fund. Any opinions, findings, conclusions, or M recommendations expressed in this material are CM those of the authors and should not be attributed to their employers or funding sources.

References

1. D. Barrett, “One Surveillance Cam- era for Every 11 People in Britain, Says (b) CCTV Survey,” The Telegraph, 10 July 2013; www.telegraph.co.uk/technology/10172298/One-surveillance-cam- Figure 6. Cloudlet placement: (a) smaller edge cloudlets (C ) versus (b) larger era-for-every-11-people-in-Britain-says- E CCTV-survey.html. metropolitan area network (MAN) cloudlets (CM). Note that IE and IM are the ingress bandwidth from the edge and MAN, respectively. 2. H. Haddon, “N.J. Mandates Cameras for Its New Police Cars,” Wall Street J., 11 Sept. 2014; www.wsj.com/articles/n- j-mandates-cameras-for-its-new-police- cars-1410488158. and time. As the density of high-data- retention period) for on-demand dena- 3. L. Westphal, “The In-Car Cam- rate sensors in the IoT grows, it will turing and big data processing. Finally, era: Value and Impact,” The Police become increasingly difficult to sus- GigaSight supports interactive, content- Chief, vol. 71, no. 8, 2004; www.poli- tain the practice of shipping all cap- based time-sensitive searches that the cechiefmagazine.org/magazine/index. cfm?fuseaction=display&article_id=358. tured data to the cloud for process- indexer didn’t anticipate. ing. The decentralized and federated Also, although we’ve focused on 4. A. Davies, “Here’s Why So Many Crazy architecture of GigaSight, using VMs video cameras as sensors here, any Russian Car Crashes Are Caught on Cam- era,” Business Insider, Dec. 2012; www. on cloudlets for flexibility and isola- data type in the IoT that has a high businessinsider.com/why-russian-drivers- tion, offers a scalable approach to data data rate can potentially benefit from have-dash-cams-2012-12. collection. Sampling and denaturing such a global repository. For exam- 5. Implementing a Body-Worn Cam- data immediately after capture enables ple, each GE jet aircraft engine gener- era Program: Recommendations and owner-specific lowering of fidelity to ates 1 Tbyte of sensor data every 24 Lessons Learned, Office of Com- preserve privacy. Performing edge an- hours.16 The airframe of the Boeing munity Oriented Policing Services, 2014; www.justice.gov/iso/opa/ alytics (such as indexing) in near real 787 generates half a Tbyte of sensor resources/472014912134715246869.pdf. time on freshly denatured data greatly data on every flight.17 Modern au- improves the time-to-value metric of tomobiles—especially the emerging 6. S. Banerjee and D.O. Wu, “Final Report from the NSF Workshop on Future Direc- this data. The raw data at full fidelity is self-driving family of cars—generate tions in Wireless Networking,” National still available (during the finite storage comparable amounts of sensor data. Science Foundation, Nov. 2013.

30 PERVASIVE computing www.computer.org/pervasive the Authors

Mahadev Satyanarayanan is the Carnegie Zhuo Chen is a PhD candidate in the Computer Group Professor of Computer Science at Carnegie Science Department at Carnegie Mellon University, Mellon University. He is an experimental com- working with Mahadev Satyanarayanan. His main re- puter scientist whose many pioneering contribu- search interest lies in distributed systems, mobile com- tions to distributed systems and mobile comput- puting, and the application of computer vision in such ing are in widespread use today. Satyanarayanan contexts. Specifically, he explores how the effortless received his PhD in computer science from Carn- video capture of smart glasses, such as Google Glass, egie Mellon University. He is a Fellow of the ACM can benefit people with the cloudlet infrastructure. and IEEE. Contact him at [email protected]. Chen received his BE electronic engineering from Tsin- ghua University. Contact him at [email protected]. Pieter Simoens is an assistant professor at Kiryong Ha is a PhD student at Carnegie Mellon Ghent University, Belgium. His research in- University. His research interests include mobile terests include distributed software systems, computing, virtualization, and cloud computing. Ha mobile cloud computing, and edge computing. received his MS from KAIST. He is currently work- Simeons received his PhD in computer science ing on a cloudlet project about the convergence of from Ghent University. Contact him at pieter. mobile and cloud computing. Contact him at krha@ [email protected]. cmu.edu.

Yu Xiao is currently a postdoc researcher at Wenlu Hu is a PhD student in the Computer Sci- the department of computer science and engi- ence Department at Carnegie Mellon University. neering, Aalto University. Her current research She is advised by Mahadev Satyanarayanan, and interests include energy-efficient wireless net- her research interests are in distributed and mobile working, crowd-sensing, and mobile cloud systems. Hu received her BS in electronic engineer- computing. Xiao received her PhD in computer ing from Tsinghua University. Contact her at wenlu@ science from Aalto University. Contact her at cmu.edu. [email protected].

Padmanabhan Pillai is a senior research scien- Brandon Amos is a computer science PhD student tist at Intel Labs. His current research investi- at Carnegie Mellon University. His research interests gates the use of cloud infrastructure close to the include distributed and mobile systems. Amos re- end user to support rich, immersive, latency- ceived his BS in computer science from Virginia Tech. sensitive applications on mobile devices. He Contact him at [email protected]. is more broadly interested in mobile and low- power systems and interactive applications of computer vision. Pillai received his PhD in computer science from University of Michigan. He is a member of IEEE and the ACM. Contact him at [email protected].

7. M. Satyanarayanan, “Mobile Computing: 11. P. Samarati and L. Sweeney, Protect- 2010, pp. 37–42; http://doi.acm. The Next Decade,” SIGMOBILE Mobile ing Privacy when Disclosing Informa- org/10.1145/1734583.1734593. Computing and Comm. Rev., vol. 15, no. tion: k-Anonymity and Its Enforcement 2, 2011, pp. 2–10; http://dl.acm.org/cita- through Generalization and Suppression, 15. P. Gilbert et al., “Toward Trustworthy tion.cfm?id=2016600. tech. report SRI-CSL-98-04, SRI Inst., Mobile Sensing,” Proc. 11th Workshop 1998. on Mobile Computing Systems & Appli- 8. M. Satyanarayanan et al., “The Case for cations, 2010, pp. 31–36; http://doi.acm. VM-Based Cloudlets in Mobile Com- 12. J. Shotton, M. Johnson, and R. Cipolla, org/10.1145/1734583.1734592. puting,” IEEE Pervasive Computing, “Semantic Texton Forests for Image Cat- vol. 8, no. 4, 2009, pp. 14–23; http:// egorization and Segmentation,” Proc. 16. J. Gertner, “Behind GE’s Vision for the ieeexplore.ieee.org/xpl/articleDetails. IEEE Conf. Computer Vision and Pat- Industrial Internet of Things,” Fast Com- jsp?arnumber=5280678. tern Recognition, 2008, pp. 1–8. pany, July/Aug. 2014; www.fastcompany. com/3031272/can-jeff-immelt-really- 9. P. Simoens et al., “Scalable Crowd-Sourc- 13. L. Huston et al., “Diamond: A Storage make-the-world-1-better. ing of Video from Mobile Devices,” Proc. Architecture for Early Discard in Interac- 11th Int’l Conf. Mobile Systems, Appli- tive Search,” Proc. 3rd USENIX Conf. 17. M. Finnegan, “Boeing 787s to Create cations, and Services (MobiSys), 2013, File and Storage Technologies, 2004; Half a Terabyte of Data per Flight Says pp. 139–152; http://dl.acm.org/citation. www.usenix.org/legacy/publications/ Virgin Atlantic,” ComputerWorld, Mar. cfm?id=2464440. library/proceedings/fast04/tech/full_ 2013; www.computerworlduk.com/news/ papers/huston/huston_html/diamond- infrastructure/3433595/boeing-787s-to- 10. J. Fan et al., “A Novel Approach for Pri- html.html. create-half-a-terabyte-of-data-per-flight- vacy-Preserving Video Sharing,” Proc. says-virgin-atlantic. 14th ACM Int’l Conf. Information and 14. S. Saroiu and A. Wolman, “I Am Knowledge Management, 2005, pp. a Sensor, and I Approve This Mes- Selected CS articles and columns 609–616; http://dl.acm.org/citation. sage,” Proc. 11th Workshop on Mobile are also available for free at cfm?doid=1099554.1099711. Computing Systems & Applications, http://ComputingNow.computer.org.

April–june 2015 PERVASIVE computing 31