We propose CARLA-FLMon, which can monitor the progress of running federated learning (FL) training in the open-source autonomous driving simulation software, CARLA. The purpose of CARLA-FLMon is to visually present the status and results of federated learning training, and to provide an extensible FL training environment with which FL training can be performed repeatedly with updated learning strategies through analysis. With CARLA-FLMon, we can determine what factors have positive or negative influences on learning by visualizing training data. Then, we can optimize the parameters of the FL training model to improve the accuracy of FL. With preliminary experiments of CARLA-FLMon on lane recognition, we demonstrate that CARLA-FLmon can increase the overall accuracy from 80.33% to 93.82% by identifying lowly-contributing clients and excluding them.

As cryptocurrency markets such as Bitcoin and Ethereum increase worldwide, cryptocurrency exchanges, service providers, and governments need to monitor and analyze cryptocurrency transactions. There are a few studies to analyze Bitcoin data in open-source projects. However, under massive data volume in the blockchain structure, we still meet the challenges of data processing or analysis speed and efficient data structure. For this purpose, this paper introduces BitSQL, which minimizes significantly database construction time and analysis performance using well-known RDBMS. BitSQL maintains the compact database due to a data mapping technique and a transaction split technique for Bitcoin analysis and it achieves the fast response time due to the index of the database. In addition, as BitSQL is based on the well-known RDBMS, we can easily support web integration or visualization tools. Through experiments, we demonstrate that the data mapping technique can decrease the database size by 95% compared to the simple Bitcoin data storing method. We also show that BitSQL can perform easily analysis regarding balances, transactions, and Bitcoin address clustering heuristics through SQL query statements. This study contributes to the scalable Bitcoin analysis software construction method for analyzing and tracking cryptocurrency blockchains and creating services.

Mobile app performance measurement is important to improve user experience by optimizing mobile apps. For this purpose, many mobile app testing methodologies. However, most of mobile app testing software depend on operating systems and binary execution files. In this work, we present a robot-based automatic mobile app speed measurement method. Our tool automates speed testing of mobile apps on either Android or Apple iOS. In particular, we employ a robot arm to simulate human behavior to generate user input events such as touch and swipe. For UI automation with a robot arm, we have devised UI automation by random or clickable component detection methods. We have adopted a web speed index to the mobile app to reflect user experience of performance. From experiments with our robot arm and automation software of generating user events and computing speed index, we have classified fast or slow mobile apps under the speed index metric.

Smart speakers like Amazon Echo or Google Home are popular because they provides intelligent virtual assistant services. However, smart speakers should be connected through the Internet to the voice server in the cloud that utilizes the powerful computing resources, which may cause the additional latency. In this paper, we propose a content prefetching and local control method to improve the response time of voice commands to smart speakers. Our proposal consists of local user intent detection on the device, content prefetching, and local text-to-speech processing. We select the popular voice commands regarding weather, music, news and time services, and implement a smart speaker prototype that can execute user intent detection at the device and play the prefetched contents with Google API. We compare our speaker prototype with five smart speakers by Amazon, Google, Kakao, Naver, and SKT. From five representative voice commands, we show that our local control of prefetched contents significantly improves the response time. For instance, in music and news services, the response time is reduced from 4.88 seconds to 0.11 seconds and 8.54 seconds to 0.13 seconds, respectively.

High-frequency sound beacon is used for various purposes such as advertising, location detecting, and tracking devices. Due to smartphones with speakers and microphones capable of handle high frequency, anyone can use ultrasonic beacon today. Google's Nearby Messages API or Starbucks location beacons are high-frequency data-over-sound applications that are popular to developers and users. However, these high-frequency beacons are not suited for delivering text data because they are prone to the interference of background sound and it heavily depends on the performance of the Digital-to-Analog converter (DAC) and the Analog-to-Digital converter (ADC). In this paper, we propose a method to broadcast data using data sequence control and error correction algorithm for high-frequency sound data communication and its application for an attendance book. The client receives the broadcasted JSON data encoded in UTF-8, processes the data, and reports to the server to identify the attendee location. Experiments have shown that at least 5 smartphones can process at the same time and the operation is guaranteed to 3-meter distance from the speaker. This study shows that data communication using high-frequency sound can be applied to a wide range of applications when sequence control and error correction are properly performed.

Voice-controlled smart speakers are popular due to Amazon Echo and Google Home. Though many smart speakers have appeared in the market, we do not know the exact performance of smart speakers. In particular, when we call a smart speaker to issue a voice command, we recognize that the smart speaker is ready when it turns its LED lamps on. The wake-up response time is important for improving user experience. Therefore, we have to measure the wake-up response time, which requires the video analysis of smart speaker LED's status for the voice command. In this work, we present a smart speaker wake-up response time measurement method with a camera. We record video files with camera while testing a smart speaker and analyze image similarity to find a wake-up event. For this purpose, we use the Google Inception-v3 model to improve the accuracy of detecting the image change of a small area.

Round Trip Time (RTT) has a decisive impact on Quality of Experience (QoE) and is therefore used as a key indicator in performance analysis. The location of the remote server is known as the most important indicator of the RTT, which is the basis of the server installation such as CDN and Proxy. However, in modern networks, private networks such as Wi-Fi and home routers are installed at the edge of the network, making it difficult for an ISP to measure its own RTT. In this study, we analyzed the relationship between server location and RTT to cover a private network like a home router. The correlation coefficient between the RTT and the server location over 100 ms is 0.33, which indicates that the distance to the server is the main factor that influences QoE when communicating with specific servers. We have gathered participants for the experiment to collect the data and have analyzed the characteristics of the group. Our research contributes to the field of analyzing the actual performance of the client with the consideration of the private network.

An IP geolocation database provides location information for each IP address block. Though free and commercial IP geolocation databases are available, their accuracy and coverage are not clearly known for non US countries. In this paper, for the wider and more accurate IP geolocation service in Korea, we present a crowd-sourcing IP geolocation database build-up method by analyzing messages of an online used market, Ruliweb that reveals IP addresses and location information. From 195,937 posts on the Ruliweb website over 29 months, we can construct a Korean IP geolocation database that achieves the district-level granularity and outperforms commercial services in the accuracy. Though every online website does not provide IP address information, our crowd-sourcing methodology is useful for IP geolocation database management in Internet or mobile services.

Tor is a commonly used anonymous network that provides the hidden services. As the number of hidden services using Tor's anonymous network has been steadily increasing every year, so does the number of services that abuse Tor's anonymity. The existing research on the Tor is mainly focused on Tor's security loopholes and anonymity. As a result, how to collect and analyze the contents of Tor's hidden services is not yet in full swing. In addition, due to the slow access speed of the Tor browser, it is difficult to observe the dynamics of the hidden services. In this work, we present a tool that can monitor the status of hidden services for the analysis of authentic hidden service contents, and have automated our tool with the virtualization software, Docker, to improve the crawling performance. From August 12, 2017 to August 20, 2017, we collected a total of 9,176 sites and analyzed contents for 100 pages.

One of the smartphone's killer applications is the game. As the mobile game market grows drastically, we meet a lot of new smartphone games on the application stores everyday. In particular, mobile games have been equipped with social networking features, which is so called social mobile games. In general, mobile games usually require real-time transmission delay and minimum bandwidth contraints. However, troubleshooting or monitoring of mobile game traffic is not straightforward. Performance degradation during the online game is not easily detected or cannot be quickly resolved because of the complicated components of the smartphone game platform, which consists of accouting, distribution, social networking, application stores, advertisement, and related API servers. In this work, we present a measurmeent-based view of a representative social mobile game, Anipang, in Korea. From experiments, we show how Anipang works based on the social network service, Kakaotalk, and present case studies of performance anomalies.

Since peer-to-peer (P2P) applications like BitTorrent are prevalent on the Internet, the amount of Internet traffic is explosively growing. The traffic caused by P2P applications to share large files have a considerable impact on the network operation, because P2P applications significantly consume network resources and often saturate the capacity of broadband access, inter-domain, or international links. Currently, in order to optimize traffic, IETF Application-Layer Traffic Optimization (ALTO) WG standardizes a traffic localization protocol that provides location and ranking information to peers, that will be useful for reducing overall traffic load by connecting nearest peers instead of random ones. In this work, we verify the effectiveness of ALTO protocols from intensive simulations on a scaled-down commercial network topology. From simulation results, we could observe that the total amount of traffic, the load of inter-domain and international links, and the user download time could be remarkably diminished.

To devise a large-scale network management system, we should consider following issues: heterogeneity of network devices, interoperation between multiple network domains, scal- ability and availability of the management system, and the cost of developing and deploying the system. Traditional network management software does not provide a cost-effective solution for the above issues because they are usually designed for a specific network domain composed of a limited set of network devices, and they have a complex multi-tiered architecture that usually requires a separate failover scheme and development strategy for each tier at the additional costs of development, deployment, and operation. In this paper, we propose a novel network management architecture, CORD, based on consistent hashing. CORD achieves scalable load balancing and sub-50ms failover performance against a node failure with a one-tired architecture. Using XML, CORD provides highly-customizable southbound management interface which integrates network devices from multiple vendors with ease. With the experiments, we have shown that the standard deviation of load balancing on CORD nodes is maintained around 5% and the failover time is below 50ms on average. In addition, we have applied CORD to the optical network management system development to reduce CAPEX and OPEX.

When we reverse-engineer unknown protocols or analyze the Internet traffic, it is critical to capture complete traffic traces generated by a target application. Besides, to prove the accuracy of Internet traffic classification algorithms of the traffic monitoring system usually located in the middle of the network, it is highly required to retain traffic traces associated with the related application. Therefore, in this paper, we present an application-specific packet capturing method at end hosts, which is based on the dynamic kernel probing technique. From the experiments it is shown that the proposed method is useful for creating per-application complete traffic traces without performance degradation.

In order to support a line-speed intrusion detection, a hardware-based solution for an intrusion detection is required, because the traditional Intrusion Detection System (IDS) based on software does not provide enough performance. In this paper, we have proposed a network processor based intrusion detection system supporting up to a 10 Gbps interfaces with deep packet inspection. To achieve higher performance, we have employed a Ternary Content Addressable Memory (TCAM) and invented a parallel TCAM accessing scheme with storing all position-aware sub-patterns in TCAM. In addition, the parallel TCAM access is applicable to the multi-packet inspection and plenty of very large patterns such as virus and worm signatures. With various experimental results, we have clearly justified the proposed algorithm works well for a multi-gigabit intrusion and virus/worm detection system.

With the increasing importance of network protection from cyber threats, it is requested to develop a multi-gigabit rate pattern-matching method for protecting against malicious attacks in high-speed network. This paper devises a high-speed deep packet inspection algorithm with TCAM by using an m-byte jumping window pattern-matching scheme. The proposed algorithm significantly reduces the number of TCAM lookups per payload by m times with the marginally enlarged TCAM size which can be implemented by cascading multiple TCAMs. Due to the reduced number of TCAM lookups, we can easily achieve multi-gigabit rate for scanning the packet payload. It is shown by simulation that for the Snort rule with 2,247 patterns, our proposed algorithm supports more than 10 Gbps rate with a 9 Mbit TCAM.