Ensuring Data Integrity and Freshness in GPU-CXL Transfers With Tamper-Resistant Metadata

GPUs have become critical accelerators in applications such as scientific computing and deep learning. The demand for more robust security measures has significantly increased as their usage expands across a broader range of applications. To address this need, Trusted Execution Environments (TEEs) have been integrated into GPU systems to safeguard applications and data. However, with technologies like Compute Express Link expanding the memory capacity of GPUs, cross-memory data transfers now pose new security challenges. Existing methods attempt to optimize the performance overhead associated with these transfers by sharing security metadata. However, this shared metadata is unreliable, raising concerns about data integrity and freshness and making systems vulnerable to replay attacks. In response to these issues, this paper introduces the trusted shared security metadata concept. Leveraging the tamper-resistant of MoveCtr, the paper develops a trusted shared security metadata scheme and a secure cross-memory integrity tree, ensuring the integrity and freshness of transferred data and protecting against replay attacks. Experimental results demonstrate that this method provides strong security guarantees while maintaining performance overhead within an acceptable range.