Android Binary XV Diffing Code and Data

This paper presents XvDiff, a high-precision and high-performance cross-version diffing tool for Android binaries that integrates structural code-feature similarity with LLM-based semantic reasoning. XvDiff accurately identifies function-level changes even under heavy obfuscation, substantially outperforming state-of-the-art diffing approaches. Leveraging the diff results produced by XvDiff, we construct XvCKG, a multi-version Android code knowledge graph that unifies both intrinsic version-specific information and cross-version differential information into a single structured representation. Built upon this knowledge graph, we further develop an automated analysis framework that employs agentic retrieval-augmented generation to perform cross-version reasoning, enabling the detection of functional changes and the evolution of security properties throughout app development.