APICraft: Fuzz Driver Generation for Closed-source SDK Libraries

Fuzz drivers are needed for fuzzing libraries. A fuzz driver is a program which can execute library functions by feeding them with inputs provided by the fuzzer. In practice, fuzz drivers are written by security experts and the drivers' quality depends on the skill of their authors. To relieve manual efforts and ensure test quality, different techniques have been proposed to automatically generate fuzz drivers. However, existing techniques mostly rely on static analysis of source code, leaving the fuzz driver generation for closed-source SDK libraries an open problem. Fuzz driver generation for closed-source libraries is faced with two major challenges: 1) only limited information can be extracted from the library; 2) the semantic relations among API functions are complex yet their correctness needs to be ensured. To address these challenges, we propose APICRAFT, an automated fuzz driver generation technique. The core strategy of APICRAFT is collectcombine. First, APICRAFT leverages both static and dynamic information (headers, binaries, and traces) to collect control and data dependencies for API functions in a practical manner. Then, it uses a multi-objective genetic algorithm to combine the collected dependencies and build high-quality fuzz drivers. We implemented APICRAFT as a fuzz driver generation framework and evaluated it with five attack surfaces from the macOS SDK. In the evaluation, the fuzz drivers generated by APICRAFT demonstrate superior code coverage than the manually written ones, with an improvement of 64% on average. We further carried out a long-term fuzzing campaign with the fuzz drivers generated by APICRAFT. After around eight month's fuzzing, we've so far discovered 142 vulnerabilities with 54 assigned CVEs in macOS SDK, which can affect popular Apple products such as Safari, Messages, Preview and so on.