SBIR Phase I: Secure Image Recognition and Machine Learning Using Advanced Cryptography

这一小型企业创新研究(SBIR)第一阶段项目的更广泛影响/商业潜力将是朝着解决大数据时代访问与隐私的两难境地迈出的重要一步。使用人们的生物识别、互联网流量以及金融、医疗和基因数据可以更好地预防犯罪、定向广告和健康创新，但代价是隐私。数据也可能过于敏感，不能提供给第三方。采用这项技术的直接影响将是提高敏感图像数据的安全性，使其更容易获得有用的推断。该解决方案将把机构在现场存储敏感数据的模式转变为即使是敏感数据也在云中存储和访问的模式。随着私人外包数据分析的能力，将带来一个计算任务的市场，包括作为服务的机器学习，这将促进研究，并以更快的速度向患者和客户提供更好的结果，而不会有暴露风险。这个小型企业创新研究(SBIR)第一阶段项目将采用现有的深度神经网络模型，使用完全同态加密方案对加密图像执行图像分类。主要的挑战是减少对加密数据操作的计算开销，以使该方案在期望的精度和安全性水平上切实可行。拟议的研究和开发通过在机器学习、计算数论、近似理论和计算机科学方面的创新来解决这一挑战。拟议的研究和开发的目标是通过实现合理的安全性、准确性和服务器成本来证明安全图像识别的商业可行性。该团队将试验使用精心选择的激活函数和/或测试函数的近似来训练和测试用于图像分类的改进卷积神经网络(CNN)，同时在现有的同态加密库上构建新的功能来同态计算这些运算。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project will be a significant step towards resolving the access vs. privacy dilemma of the big data era. The use of people’s biometrics, internet traffic, and financial, medical, and genetic data can enable better crime prevention, targeted ads, and health innovation, but at the expense of privacy. Data may also be too sensitive to be given to third parties. The immediate impact of adopting this technology will be greater security for sensitive image data with easier access to useful inferences. The solution will shift the paradigm of institutions storing sensitive data onsite to one in which even sensitive data is stored and accessed in the cloud. With the capability of private outsourced data analysis will come a marketplace for computational tasks, including machine learning as a service, that will spur research and deliver better results to patients and clients faster and without risk of exposure.This Small Business Innovation Research (SBIR) Phase I project will adapt existing Deep Neural Network models to use a fully homomorphic encryption scheme to perform image classification on encrypted images. The primary challenge is to reduce the computational overhead of operations on encrypted data to make the scheme practical at desired levels of accuracy and security. The proposed research and development addresses this challenge through innovation in machine learning, computational number theory, approximation theory, and computer science. The goal of the proposed research and development is to demonstrate the commercial viability of secure image recognition by achieving a reasonable level of security, accuracy, and server cost. The team will experiment in training and testing modified convolutional neural networks (CNNs) for image classification using carefully chosen activation functions and/or approximations to the testing function, and simultaneously building onto existing homomorphic encryption libraries new functionality to compute these operations homomorphically.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.