Final Projects Summary

Automotive Anomaly Detection with Resistence to Adversarial Samples

Members: Mengyu Liu, Lin Zhang
Summary: Cyber-Physical system is a field integrate communication, controlling and computing with multiple sensors and actuators connected to the physical world. Attacks from invader and complex environment could lead to abnormal state of the system. Many previous works designed various anomaly detection algorithms to handle this problem. Recently, machine learning methods has been applied in many works but anomaly samples are extreme rare which makes the performance not good. An efficiency way is data augmentation, a few works applied Generative neural networks to achieve better performance, but none of them explain how the process related to physical world and did interpret it from a knowledge-based view. Our work proposed a novel perspective on how to understand how Generative Neural Network helped exploiting the knowledge of our collected data with different anomaly detection algorithm. Additionally, we built testbeds to simulate autonomous vehicles to see how the augmented data related to physical world.
Presentation

Comparing Model Accuracy for Differential Privacy and Machine Unlearning on Sensitive Data

Members: Vedhas S. Patkar and Joel W. Yuhas
Summary: Machine Learning models, particularly those that employ neural networks, generally need large datasets for identification. Our aim is to employ existing models that use differential privacy and machine unlearning to determine the best possible outcomes for sensitive and identifiable data, both in terms of privacy and model accuracy. We will set out to find the limitations and advantages of existing architectures, and figure out the possible changes that can be made to them to provide better outcomes
Presentation

Asynchronous Federated Learning - Literature Review

Members: David Castello
Summary: Federated learning is an emergent field of research that seeks to train performant machine learning models across a heterogeneous collection of mobile devices, each with access to private data. Despite the fact that such devices have uneven hardware capabilities and quantities of training data, and that network constraints and hardware failures can occur at any time, state of the art algorithms to conduct federated learning operate in synchronous rounds of communication. This work is a literature review of recent efforts to address this contradiction via novel protocols that relax the need to synchronize local training. Following a background study of the influences behind federated learning, six new research papers are surveyed to assess the progress of this new field towards developing robust and effective asynchronous federated learning algorithms.
Presentation

Learning a Fair Model under Privacy Constraints

Members: Cuong Tran
Summary: In this class project we study binary classification problems under three angles: accuracy, privacy and fairness. We explore how to maximize model’s accuracy given the constraints that the models’ prediction should not be discriminative towards certain protected groups. At the same time, we aim the model should be secured in term of protecting training data.
Presentation

Improved StarGAN for Generating Adversarial Samples

Members: Jindi Wu, Mu Bai
Summary: We propose a solution of adversarial samples generation based on the StarGAN, which is an image-to-image translation model and can smoothly generate a set of new images from source images. The images generated by StarGAN could be adversarial samples for attacking deep neural network classifier. Our main aim is to attack the targeted model with specific output label and pass the corresponding smooth detection simultaneously.
Presentation

Constrained Deep Learning for Fast Approximation of Scheduling Problems

Members: James Kotary
Summary: Combinatorial optimization problems in planning and scheduling are known to lack efficient solution methods,while demand for the frequent and accurate solution of such problems is ever increasing across several engineering fields. Recent developments in machine learning promise new approaches for the fast approximation of constrained optimization problems at significantly reduced time-scales. This project will explore the application of constrained deep learning to build a system that predicts accurate solutions to challenging scheduling problems, with favorable properties not reflected in the original problem structures.
Presentation

Privacy Preserving with Preference Elicitation Process

Members: Pratik Ashok Paranjape and Raman Srivastava
Summary: Preference elicitation is the process of developing a decision support system to generate recommendations for a user. Privacy preserving refers to the system of publicly sharing information about a dataset by describing the patterns of groups within the dataset without dis-closing the information about individuals in the dataset. The idea of this project is to implement differential privacy in a recommendation system.
Presentation