부산대학교 도서관

Mobile networks can connect everyone with a simple call or message and recent years have seen a dramatic increase in abuse through robocalling and messaging. This impacts everything from network optimization decisions to users being scammed. Users on any network often believe that there should be some level of intelligence that allows for the blocking and detection of abuse like robocalling, telemarketing, or scam/fraud schemes. A key failure in most approaches is the desire to maintain user privacy while collecting this kind of intelligence. We have designed a way to determine trust between two devices in a way that is driven by device behavior and not intrusive through surveys or requiring access to private contact lists. Instead, we observe encrypted network traffic to build a weighted graph by taking all traffic as inputs to create a weighted vector which in turn considers all other traffic on the network and scales to a single numeric representation of trust. This gets us to a weighted trusted graph that can be the foundation to many other important business use cases including fraud prevention, marketing, and churn. After exploring our trusted graph, a form of social graph with trusted weights, we found that our graph is hyper connected through hub nodes that have very high outdegree counts, many relationships. In other social graphs this might be referred to as influencers, but when we add the concept of trusted relationships, these might be unwanted influencers. We show that by setting the right trusted threshold, we can filter out ~80% of the relationships in our graph without losing the core social structures, suggesting that a potential super-majority of relationships are untrusted.