Serving deep learning based recommendation models (DLRM) at scale is challenging. Existing approaches rely on dedicated ANN indexing and filtering services on CPUs, suffering from non-negligible costs and missing co-design opportunities. Such inefficiency makes them difficult to support complex model architectures, such as learned similarities and multi-task retrieval. In this paper, we present SilverTorch, a model-based serving system that brings all components into one unified model. It unifies model serving by replacing standalone indexing and filtering services with model layers. We propose a model-based GPU Bloom index for feature filtering and a fused Int8 ANN kernel for nearest neighbor search. Through co-design of the ANN search and feature filtering, we reduce GPU memory usage and eliminate computation. Benefiting from this design, we scale up retrieval by introducing an OverArch scoring layer and a multi-task retrieval with a Value Model to aggregate scores. These advancements improve the retrieval accuracy and enable future studies for serving more complex models. Our evaluation on industry-scale datasets shows that SilverTorch achieves up to 23.7× higher throughput compared to the state-of-the-art approaches. We also demonstrate that SilverTorch's solution is 13.35× more cost-efficient than CPU-based solution while improving accuracy via serving more complex models.

Recent studies have proposed moving target defense (MTD) to detect false data injection (FDI) attacks in power grids. To hide the activation of MTD from attackers, a hidden MTD (HMTD) has been proposed, which keeps the system power flow after MTD unchanged. It has been proved that HMTD cannot detect all FDI attacks because of its stealthiness requirements. However, the mathematical mechanism of MTD's stealthiness has yet to be revealed. The maximum detection capability of HMTD is also unclear. To address the abovementioned issues, we first analyze the maximum detection capability of HMTD based on graph theory and propose the topological condition to achieve it. Moreover, we study the essential characteristics of HMTD and find that all HMTD schemes are in a space spanned by branch parameters. We further propose a multistage HMTD (MHMTD) method to select multiple HMTD schemes in this space to maximize the detection capability. Experiments show that the MHMTD can maximize the detection capability of HMTD in all test systems with high stealthy probability.

In recent studies, moving target defense (MTD) has been applied to detect false data injection (FDI) attacks using distributed flexible ac transmission system (D-FACTS) devices. However, the inherent conflict between the security goals of MTD (i.e., detecting FDI attacks) and the economic goals of D-FACTS devices (i.e., reducing power losses) would impede the application of MTD in real systems. Moreover, the detection capabilities of existing MTDs are often insufficient. This article proposes a multistage MTD (MMTD) approach to resolve these two issues by adding a group of designed security-oriented schemes before D-FACTS’ economy-oriented scheme to detect FDI attacks. We keep these security-oriented schemes for a very short time interval and then revert to the economy-oriented scheme for the remaining time to ensure the economic requirements. We prove that a designed MMTD can significantly improve the detection capability compared to existing one-stage MTDs. We find the supremum of MMTD’s detection capability and study its relationship with system topology and D-FACTS deployment. Meanwhile, a greedy algorithm is proposed to search the MMTD strategy to reach this supremum. Simulation results show that the proposed MMTD can achieve the supremum against FDI attacks while outperforming current MTD strategies on economic indicators.