Verifying the Robustness of KNNs against Data-Poisoning Attacks

In this work, we present a method for proving data-poisoning robustness of the widely used K-nearest neighbors (KNN) algorithm. Our method can soundly over-approximate the KNN behaviors during both learning (parameter tuning) and prediction phases, while covering all possible scenarios in which up to n training data elements are maliciously injected. We also develop optimizations to prune the search space while maintaining accuracy.

Constraint-Based Precomputation

Energy harvesting is an environment-friendly technology that converts ambient energy to electricity to power computing devices in the Internet of Things. Due to the unstable energy source, software on these devices must deal with intermittent power supply and mismatched demand. We proposed a program analysis based method for optimizing energy utilization automatically. Œ

Our idea is to identify precomputation demands, i.e., the ones that may be performed ahead of time, and reorder them to match the intermittent power supply. We develop a sound static analysis to identify precomputation opportunities, an SMT solver based method to optimize the precomputation policy, and a semantic-preserving transformation to generate the optimized program.

Privacy-Preserving Image Trading through Crowdsourcing

I led a five-member team. We proposed and implemented a novel ownership and privacy-aware image trading system based on crowdsourcing. We designed a novel image evaluation and selection method based on CNN(ConvolutionalNeuralNetwork)model. We minimized computation and communication overhead in both servers and clients sides. Our evaluation and experiments show high quality of our selected datasets and high efficiency of our methods. This work was accepted by INFOCOM 2018 (first-student author, second author).

Optimized a 2-path Shortest Algorithm on CPU/GPU Heterogeneous Platform

I implemented and optimized a 2-D shortest path algorithm on CPU/GPU heterogeneous platform. This project is for outstanding undergraduates, which is sponered by Mitacs Globalink Research Intern Grant from Canada Mitacs.