SONAR

Social networks engender implicit trust of knowing the person behind the information, so when posts are presented in our social media timelines, they have the potential to persuade us and change our opinions. For other users, such as intelligence analysts, while information shared on social networks is unlikely to take the place of vetted intelligence, the presence of bots can make it difficult to focus on potentially useful signals being shared on Twitter, and also to focus our own engagement campaigns on real users. To regain situational awareness of one’s actual (human) social network, we need a tool that can accurately detect a wide range of malicious or misleading accounts (bots, trolls, foreign actors, etc.), assess exposure in terms of our level of engagement with the communities most compromised by such accounts, and provide a means to contain, mute, or block problematic parts of our network.  To address this need, Systems & Technology Research (STR) and our teammate, Prof. Tauhid Zaman from Yale University, propose to develop SONAR (social vulnerability assessor), a data-driven app for understanding vulnerabilities in a Twitter social network. SONAR combines detection of malicious accounts (e.g., bots, trolls, bad actors) in a user’s social network communities with that user’s level of online engagement to compute a vulnerability index that captures a user’s exposure to outside influence. When a user logs into the SONAR app using Twitter credentials, the system retrieves social network data (e.g., friends/followers and recent tweets) to compute a set of multi-modal features. These features allow SONAR to discover communities in the network and detects malicious accounts among them. SONAR provides each community with a semantic label and evaluates both its threat level and the user’s social engagement with compromised accounts in the network using a novel opinion dynamics model. Lastly, SONAR computes a vulnerability score and displays all the inferred information in the SONAR app. The app will issue alerts of detected vulnerabilities and threats, and allow users to control their social network by muting or blocking specific members or whole communities. Our work combines the team’s state-of-the-art research and experience in social network science, probabilistic modeling, machine learning and natural language processing with a flexible app development framework to deliver an effective prototype capability under limited time and budget constraints. We note that the technology and deliverables developed under this program have dual use applications and should be determined to have both civil and military applications.