With the growing popularity of emerging technologies, the prevalence of digital systems is more than ever. Security, however, has still lagged behind, as evidenced by the increasing rate of recent attacks. Oftentimes, cyber-attacks are initiated by running a malicious code or by exploiting vulnerabilities in the underlying software. To mitigate such alarming threats, analyzing software binary code (a.k.a. binary analysis) has been known as a promising
solution. This thesis answers the following research question: how to automatically fingerprint a cross-architecture code with optimization and obfuscation by attributing compiler provenance, identifying library functions, and detecting vulnerable functions? Specifically, it first analyzes the syntax, structure, and semantic of functions to extract compiler provenance in cross-complied binaries. Second, it introduces a single robust function signature
based on heterogeneous features to solve library function identification problem. Third, it overcomes vulnerable function detection problem through a multi-stage fuzzy matching approach on firmware images. Finally, it addresses vulnerability detection problem in cross-architecture obfuscated binaries and firmware images through a neural machine translation-based approach. This thesis advances the state-of-the-art by improving the accuracy, scalability, and efficiency of binary code analysis. All of the proposed approaches are implemented as a prototype system and their performance are evaluated with real data.