In the next decade, program analysis will be facing substantial and unsurmountable issues. Analysis of third-party programs will be hard to impossible, as programmers can deliberately make their programs unanalyzable, posing major risks for security assessment. The issues are due to fundamental limitations of code analysis (e.g., the halting problem) and execution analysis (i.e., incompleteness), which cannot be overcome by better methods. In real-world settings, this effectively disables our ability to predict program properties. On top, all techniques need detailed formal specifications of expected and unexpected properties, which hardly exist. To overcome these problems, we suggest a strong research focus in the following areas, which so far are not part of the program analysis portfolio: Specification mining to infer what a program “normally” does; Test generation to systematically explore behavior for specification mining; and Sandboxing to prohibit abnormal behavior at runtime. In early experiments, the combination of the three has shown promising results: Sandbox mining constrains a program to the behavior found during testing, thwarting common malware, attack, and backdoor strategies.
分享
引用
