Modern Java compilers, such as Just-In-Time (JIT) and Ahead-OfTime (AOT) compilers, often infer and analyze complex inter-class structures to perform program optimizations while preserving semantic correctness. However, incorrect inference of these structures during compilation can lead to critical bugs, as observed in production compilers like HotSpot and R8. Despite their significance, existing Java fuzzing tools fail to adequately explore inter-class relationships, focusing instead on simpler intra-class or intra-method constructs. To bridge this gap, we present InterFuzz, the first fuzzing framework designed to systematically generate Java test cases with complex inter-class structures. InterFuzz introduces a novel concept of Heterogeneous Program Graph (HPG) to abstract and manipulate inter-class relationships. It then employs Inter-Class Mutators to construct intricate interactions, and utilizes Graph Complexity to guide test generation toward high-diversity code. Our evaluation shows that InterFuzz effectively uncovers compiler bugs that elude traditional fuzzers, having discovered 24 new bugs across HotSpot, ART, and R8, with 20 confirmed by developers—16 of which hinge on intricate inter-class structures.