웹 보안 및 개인정보 보호 연구실

We investigate various security and privacy threats in recent machine learning models and implement tools that assess the robustness of given ML models by conducting penetration security testing. The paper presents the design and implementation of FraudDetective, a dynamic testing frame- work that identifies ad fraud activities. FraudDetective computes a full stack trace from an observed ad fraud activity to a user event by connecting fragmented multiple stack traces, thus generating the causal relationships between user inputs and the observed fraudulent activity. We revised an Android Open Source Project (AOSP) to emit detected ad fraud activities along with their full stack traces, which help pinpoint the app modules responsible for the observed fraud activities. We evaluate FraudDetective on 48,172 apps from Google Play Store. FraudDetective reports that 74 apps are responsible for 34,453 ad fraud activities and find that 98.6% of the fraudulent behaviors originate from embedded third-party ad libraries. JavaScript (JS) engine vulnerabilities pose significant security threats affecting billions of web browsers. In this project, we present Montage, the first NNLM-guided fuzzer for finding JS engine vulnerabilities. The key aspect of our technique is to transform a JS abstract syntax tree (AST) into a sequence of AST subtrees that can directly train prevailing NNLMs. We demonstrate that Montage is capable of generating valid JS tests, and show that it outperforms previous studies in terms of finding vulnerabilities. Montage found 37 real-world bugs, including three CVEs, in the latest JS engines, demonstrating its efficacy in finding JS engine bugs. We investigate the design architecture of EOS.IO. Based on this investigation, we introduce four attacks whose root caus