With increasing awareness of information security, data privacy gains more attention in IoT applications. To effectively conserve network bandwidth and storage space, encrypted data deduplication is widely adopted in IoT systems. Therefore, deduplication of encrypted data becomes a booming research area. To balance data privacy and deduplication efficiency, most existing schemes rely on trusted third parties (TTPs). In this study, we propose a secure and efficient scheme named adaptive dynamic grained deduplication (ADG-Dedup) for IoT applications. Without the involvement of any online TTP, the deduplication efficiency can be significantly improved. For data with high privacy, the scheme also provides semantic security. Encryption keys are protected by a threshold secret sharing mechanism, so key share controller (KSCs) are introduced. Efficiency analysis demonstrates the advantages of our scheme over others, and security analysis shows its robustness and resilience against malicious attacks.