With the advance of wearable devices, an IoT (Internet of Things) gateway should support the efficient forwarding, processing, and provisioning of the streaming data obtained from them. We propose a semantically enriched IoT gateway working on a smartphone for wearable devices. The proposed gateway supports (a) an efficient consolidation and compression of heterogeneous streaming sensor data, (b) privacy preserving data processing by semantic reasoning, and (c) a selective data sharing through their social networks according to a degree of sensitivity. Experimental results with the prototype implementation show the potential of the proposed IoT gateway for novel user experiences based on semantic web technologies.
In the landscape of the Internet of Things (IoT), various physical objects such as
wearable devices may be involved with their owners’ social networks. Thus, wearable
devices can be part of human social networks and even build their own social
networks. This brings a new paradigm called Social Internet of Things (SIoT) [
However, current IoT gateways don't provide semantic information related to
sensor data which are distributed among various wearable devices. The data streams
obtained from wearable devices containing privacy sensitive information such as
heartbeat rate can be published and shared through the Web. They may cause an
unintended exposure of private data. Thus, gateways should ensure the
privacypreservation of sensitive data [
As shown in Fig. 1, the proposed architecture is divided into three main layers: • Data collecting layer (DCL): collects raw sensor data from wearable devices and consolidates/integrates raw data. • Data abstraction layer (DAL): transforms the collected raw data into semantic data based on a domain ontology. • Data sharing layer (DSL): shares the inferred semantic data with SNS groups depending on the sensitivity degree of data and the social distance from the groups.
SNS
...
Inferred Smartphone semantic Personalized IoT Gateway data Raw sensor data sceoCnnossloloerlciddtOaDa&nttoeatmolaoigny&aSnrenemaostaaontntioiincngTSrtioprleeshDaartiang Data col ecting Data abstraction Data sharing layer (DCL) layer (DAL) layer (DSL)
WDeevaircaeb1leWearab..le. DeviceWDseevaircaebnle Since wearable devices collect heterogeneous sensor data continuously in real-time, an efficient mechanism of storing and processing streaming data is very important; especially considering resource-constrained characteristics of mobile devices. To avoid processing a large amount of data on a smartphone itself, the proposed gateway transforms and merges the raw data in consideration of the time, location and sensor type, as shown in Fig. 2. In the first phase, we split streaming data into subsets with the same location (e.g., latitude and longitude). In the second phase once again, the streaming data is split into groups based on sensor types. Finally, we merge subsets into a single record that contains statistics information such as min, max, and avg. As a large volume of streaming sensor data is accumulated continuously, an efficient data compression is an essential for resource-constrained mobile devices. Through DCL, the amount of data to be processed is reduced. However, the data representations should be more machine-interpretable. Thus, semantic annotation and reasoning on the raw data based on the proposed domain ontology (as depicted in Fig. 3) are required to draw an upper-level knowledge, in which privacy sensitive part is cut out. When semantic annotation is finished, rule-based reasoning is applied to the newly generated semantic data. We use AndroJena1 as a reasoning engine. To perform rule-based reasoning, rules are defined in advance. Decision rules are defined in order to make upper-level conclusions, as shown in Table 1. 1 https://code.google.com/p/androjena/ xsd:string
id Thing Account
hasOwner xsd:string dataType Data hasTime Time
determine hasLoc hasValue xsd:tfimroem to isEmergency Status Location Value minVal iissNIrorremguallaitry … latitude longitude avmearaxgVeaVlal xsd:boolean geo:lat geo:long xsd:double rdf:ID rdf:ID rdf:ID owl:Class owl:ObjectProperty owl:DataTypePropertyowl:subClassOf Fig. 3. A part of domain ontology 2.3
In DSL, the upper-level knowledge is shared with SNS groups depending on the
sensitivity degree in terms of privacy and social distance. The social distance is the
degree to which users are willing to associate with whom have different social
characteristics. Much research about the correlation between the social distance and the data
sensitivity reveal that they are in inverse proportion [
In order to evaluate the performance, scalability and feasibility, we have implemented the proposed IoT gateway on a smartphone. All experiments were conducted on Samsung Galaxy Note3 and Samsung Gear2. The actual data sets are provided by a smartphone and a wearable device. To enable selective data sharing, we classified the sensed data according to the sensitivity in terms of privacy. Table 2 shows the data type and classification (L1, L2, L3 and L4) used in our experiments. 3.1
In order to evaluate the proposed method of processing streaming data, we collected more than 1,500 KB of data during 5 hours (18,000 records per each sensor type). Furthermore, experiments were conducted 6 times by changing range from 2m to 60m to evaluate our approach with different conditions. It means that the moving range can be used as a threshold value to split data set into groups based on locations. We measured the sizes of compressed data and compared them with those of original data. Fig. 4 shows the sizes of consolidated data. From the result, we confirmed that a user's low-mobility pattern leads to a better data compression performance. However, in case where the moving range is too short (2~5m), the consolidated data size is similar to the original one. The reason for this is that sensors generate data every second with different locations. As a result, records are not merged.
We performed two kinds of experiments in the process of reasoning: 1) the memory consumption and 2) the processing time for different amount of input data. Fig. 5 shows the results in terms of 1) and 2). By increasing the amount of data step by step, we checked the changes of the memory usage. We found that there was no relation between the memory usage and the amount of data set. This is because the memory usage is mainly influenced by Android OS, which manages memory itself such as garbage collection. Also we measured the execution time for semantic annotation and reasoning. As shown in the result, for all the data sets, the processing times consumed by reasoning tasks were yielded as 47.2s, 77.9s, 172.6s and 449.9s, respectively. 3.3
We also demonstrated the use of the proposed gateway through a real SNS scenario on Facebook. As shown in Fig. 6, a user’s privacy information (as shown in the red box), which may indicate a user’s health status indirectly, was allowed only to her friends. Non-private information (as shown in the blue box) was allowed to the public. The experimental result shows that 1) raw stream data were integrated and translated correctly into our semantic model 2) the target audience of the post was selectively determined by the degree of sensitivity as we defined.
Original
Consolidated used memory (MB) time(s) 500 450 ) 400 c
see 350 ( 122350500000 iitsscgenm 51000 rp
o 0 1600 1400 )1200 (KB1000 ize800 s ta 600 ad400 200 0 2 5 10 20 40 60
moving range(meter)
This research was supported by the National Research Foundation of Korea (NRF) funded by Korea government (2013R1A2A2A01016327) and Samsung Electronics Co., Ltd.