Page 74 - Undergraduate Research Opportunities Program Proceedings 2017-18

School of Engineering
Department of Computer Science & Engineering

Photo and Document Digital Forensics

Supervisor: IEONG Sze Chung Ricci / CSE
Student: ANANNA Shekh Shermin Jahan / Course: UROP1100, Spring
COGBM

In an era when image manipulation is becoming more and more common, methods to identify falsified images have
become a major topic of interest. Error Level Analysis is one such tool that can identify differences in compression level
within a single image. In my research, I looked at how Error Level Analysis is implemented and used it on modified
images to understand the extent to which this tool can identify edits in images. Changes such as blurring, sharpening,
cloning and splicing are easily detectable but changes in color, saturation, brightness and contrast are not detectable.

Deep Coding: Generating Source Code Using Deep Learning Techniques

Supervisor: KIM Sung Hun / CSE
Student: CHEN Ziqi / CPEG Course: UROP1100, Fall

Source code generation has been a hot topic as it helps people to achieve repetitive and heavy programming tasks and
enhance work efficiency. In this project, I studied state-of-the-art deep learning models for source code generation, and
implemented a neural encoder-decoder model with attention mechanism to generate general-purpose code from
natural language inputs, proposed in a paper [1]. More specifically, the model encodes natural language inputs using bi-
directional LSTM, and decodes into source code by spanning an abstract syntax tree and generating tokens at tree
terminals. I implemented the model using Tensor flow framework, experimented it on two datasets and obtained similar
results as in the literature.

Supervisor: KIM Sung Hun / CSE
Student: CHUNG Jihoon / COMP Course: UROP1100, Fall

Voice conversion is a technique to aim to modify a speaker’s voice characteristics in order to sound like someone else. In
this project, we focuses on extracting data from wave files, processing the data and, finally finding the spectral mapping
component of voice conversion. We tried to achieve some of the current state of the art result which is implemented by
a deep bidirectional long short term memory recurrent neural network based voice conversion system, and also tries to
apply sequence to sequence model with attention. We were not able to make the attention layer to have meaningful
result to the total model. We believe we can improve our model with regularization.

Supervisor: KIM Sung Hun / CSE
Student: DEL MUNDO Gian Miguel Sero / COMP Course: UROP1100, Fall

Autonomous driving has recently become a hot topic in computer science, but state-of-the-art approaches have not
been able to account for non-deterministic scenarios, semantic information, interactions with other agents such as
pedestrians, or long-term possibilities. The project explores a proposal by Lee et al. [1] called the Deep Stochastic IOC
RNN Encoder decoder network (DESIRE) to overcome the aforementioned limitations. It consists of two modules,
namely the Sample Generation Module (SGM) and the Ranking & Refinement Module (RRM). The SGM produces a
diverse set of future trajectory predictions based on the past trajectory of the car, and the RRM scores these predictions
based on semantic information retrieved by a convolution neural network (CNN) as well as interactions with other
agents.

64