AI-as-a-Service has emerged as an important trend for supporting the growth of the digital economy. Digital service providers make use of their vast amount of customer data to train AI models (such as image recognitio...
详细信息
AI-as-a-Service has emerged as an important trend for supporting the growth of the digital economy. Digital service providers make use of their vast amount of customer data to train AI models (such as image recognition, financial modelling and pandemic modelling etc) and offer them as a service on the cloud. While there are convincing advantages for using such third-party models, the fact that model users are required to upload their data to the cloud is bound to raise serious privacy concerns, especially in the face of increasingly stringent privacy regulations and legislation. To promote the adoption of AI-as-a-Service while addressing privacy issues, we propose a practical approach for constructing privacy-enhanced neural networks by designing an efficient implementation of fully homomorphic encryption. With this approach, an existing neural network can be converted to process FHE-encrypted data and produce encrypted output which are only accessible by the model users, and more importantly, within an operationally acceptable time (e.g., within 1 s for facial recognition in typical border control systems). Experimental results show that in many practical tasks such as facial recognition, text classification and so on, we obtained the state-of-the-art inference accuracy in less than one second on a 16 cores CPU.
Since the efficiency of current automatic calibration system for high precision pointer meter is limited, this paper developed a new meter calibration system based on machine vision which applied a calibration method ...
详细信息
ISBN:
(纸本)9781467377232
Since the efficiency of current automatic calibration system for high precision pointer meter is limited, this paper developed a new meter calibration system based on machine vision which applied a calibration method based on look-uptable (LT) algorithm. Advanced machine vision and image processing algorithms are employed to recognize the readings automatically. Sufficient experimental results demonstrate that the system can recognize the reading of the precise meter automatically and accurately, and significantly improve the verification efficiency.
In this paper, we focused on the retrieval of the LAI in an alpine wetland located in western part of China in late August and early July 2011. A two-layer canopy reflectance model (ACRM) was used to establish the rel...
详细信息
In this paper, we focused on the retrieval of the LAI in an alpine wetland located in western part of China in late August and early July 2011. A two-layer canopy reflectance model (ACRM) was used to establish the relationships between the LAI and the reflectance of near-infrared (NIR) and red (RED) wavebands. The reflectance data were derived from Landsat TM LIT product and the Terra and Aqua MODIS 16-day and 8-day composite reflectance products (MOD/MYD09) at 250m resolution. Due to the lack of the information about some major input parameters for ACRM, which are sensitive to model outputs in the reflectance of NIR and RED wavebands, the inverse problem was ill-posed. To overcome this problem, a method of increasing the sensitivity of the LAI while reducing the influence of other model free parameters based on the study of free parameters' sensitivity to the ACRM outputs and the region's features was studied. The area of interest was divided into two parts using the approximately statistic normalized difference vegetation index (NDVI) value around 0.5. One part was sparse vegetation (0.1 < NDVI < 0.5), which is more sensitive to soil background effects and less sensitive to the canopy biophysical and biochemical variables. The other part was dense vegetation (0.5 <= NDVI < 1.0), which is less sensitive to soil background effects and more sensitive to plant canopies and leaf parameters. Then, the relationships of rho(nir)-LAI and rho(red)-LAI were established using a look-up table algorithm for the two parts. Furthermore, a regularization technique for fast pixel-wise retrieval was introduced to reduce the elements of LUT sets while maintaining a relatively high accuracy. The results were very promising compared to the field measured LAI values that the correlation (R-2) of the measured LAI values and retrieved LAI values reached 0.95, and the root-mean-square deviation (RMSD) was 0.33 for late August, 2011, while the R-2 reached 0.82 and RMSD was 0.25 for early Ju
暂无评论