The recent developments in the research area of secure computing and machine learning have shown a significant shift towards more efficient and scalable solutions for protecting sensitive data and models. Researchers are increasingly focusing on partitioning models and data to leverage the strengths of both Trusted Execution Environments (TEEs) and GPUs, thereby optimizing computational efficiency without compromising security. This trend is exemplified by the introduction of novel strategies that separate privacy-sensitive components from the rest of the model, allowing for more granular protection and reduced computational overhead. Additionally, there is a growing emphasis on automating the identification of security-sensitive code for TEE isolation, which streamlines the process and reduces the Trusted Computing Base (TCB). In the realm of machine learning, advancements in zero-shot learning and binary code similarity detection are pushing the boundaries of what can be achieved with limited data and varying compilation configurations. These innovations not only enhance the robustness of models against unseen classes and obfuscated code but also pave the way for more sophisticated and adaptable security measures in the face of evolving threats.

Noteworthy papers include one that introduces a novel partition before training strategy for DNN models, significantly reducing computational costs while maintaining full model protection, and another that proposes a Visual-Semantic Graph Matching Net for zero-shot learning, achieving superior performance by leveraging semantic relationships among classes.