Provenance-based Data Trustworthiness Assessment in Data Streams

Feb 17, 2010

Download: MP4 Video Size: 379.8MB  
Watch on YouTube

Abstract

This talk presents a systematic approach for estimating the trustworthiness of data items in data stream environments (such as sensor networks). The approach uses the data item provenance as well as their values. To obtain trust scores, the approach exploits a cyclic framework which well reflects the inter-dependency property: the trust scores of data items affect the trust scores of network nodes, and vice versa. The trust scores of data items are computed from their value similarity and provenance similarity. The value similarity comes from the principle that "the more similar values for the same event, the higher the trust scores," and we compute it under the assumption of normal distribution. The provenance similarity is based on the principle that "the more different provenances with similar values, the higher the trust scores," and we compute it using the tree similarity. Since new data items continuously arrive in DSMSs, we need to evolve (i.e., recompute) trust scores to reflect those new items. As evolution scheme, we propose the batch mode for computing scores (non)periodically along with the immediate mode. Experimental results show that the approach is efficient and effective in data stream environments.