Accelerator, device and application method for defensive reasoning against resistance deep learning
By designing an accelerator for adversarial deep learning defense inference, the problem that existing accelerators cannot support multiple adversarial machine learning defense algorithms is solved, and a high-performance computing platform is implemented that supports deep learning and the execution of multiple defense algorithms.
CN119918652AActive Publication Date: 2025-05-02NAT UNIV OF DEFENSE TECH
Patent Information
- Application Number
- CN202411728227.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-28
- Publication Date
- 2025-05-02
- Estimated Expiration
- 2044-11-28
Smart Images
Figure CN119918652A_ABST
Abstract
The invention discloses an accelerator and equipment for defending and reasoning against resistance deep learning and an application method.The accelerator comprises a scheduling controller, a processing core array, a core private cache, a global cache and a memory access component, and the scheduling controller is provided with an access structure for host CPU access; the scheduling controller is respectively connected with the processing core array, the core private cache and the memory access component, the core private cache is connected with the memory access component through the global cache, the core private cache is also directly connected with the memory access component, the processing core array is formed by connecting a plurality of processing core sub-arrays through the processing core connecting module, and the processing core sub-arrays are connected with the memory access component through the global cache. The processing core sub-array is composed of a plurality of computing units. The method aims at achieving optimization design of the accelerator for antagonistic deep learning defense reasoning, the deep learning function is supported, meanwhile, multiple different antagonistic machine learning defense algorithms can be widely supported, and a high-performance computing platform can be provided for a trusted artificial intelligence system.
Need to check novelty before this filing date? Find Prior Art
Citation Information
Patent Citations
Deep neural network inference method and device in edge environment
CN112862083A
Multi-dimensional parallel artificial intelligence processor architecture
CN113191491A
Matrix and vector manipulation to support machine learning reasoning and other processes
CN115964598A
Testing adversarial robustness of systems with limited access
US20200242250A1