Treasures @ UT Dallas
Not a member yet
7697 research outputs found
Sort by
Towards Securing Untrusted Deep Neural Networks
No full textDeep Neural Network (DNN) models have achieved remarkable success in various domains,
ranging from image recognition to natural language processing. However, the increasing
reliance on cloud-based services and the proliferation of machine learning applications have
raised concerns regarding the security and privacy of these models. Protecting untrusted
DNN models from malicious manipulation and exploitation has become a critical challenge.
This dissertation addresses the issue of protecting untrusted DNN models from malicious
manipulation (ie, Trojan attack) and proposes a framework to enhance their security against
Trojan (backdoor) attacks. The framework consists of multiple dimensions of defenses that
collectively aim to safeguard the integrity of the models.
First, we introduce the background of Trojan attack and the settings of each proposed method.
Specifically, we propose two kinds of defense approaches against current Trojan attacks, one is
for model-level and another is for input-level. Both two proposed defense approaches are built
upon the most practical scenario, i.e., Black-box and Hard-Label scenario. The Black-box
means that the defender can not access the detailed parameters for the target model; while
the hard-label implies that the defender can only access the final prediction label for the
target DNN models. To the best knowledge, such a scenario is one of the most practical and
challenging scenarios for Trojan Defense.
To tackle the Trojan attacks, we propose two methods, i.e., AEVA and SCALE-UP. AEVA
is a novel Trojan detection approach that is implemented upon suspicious models. As for
SCALE-UP, it is an input-level Trojan defense technique, which is implemented upon the
input data during the inference phase. Both two techniques are inspired by certain intriguing
properties of DNN models and shown effective in the backdoor defense task.
Lastly, we discuss the potential adaptive attacks against our defense approaches and evaluate
their effectiveness. We find that our defense approach can still perform robustness against
potential adaptive attacks
The Contemporary Farr(ah)
No full textThis thesis is about the personal struggle and desire for peace and unity. Time-based
installations made for this thesis are built by using invisible ink, blacklight LEDs, acrylics, and
computing technologies and are the result of two- and half-year experimentation with an
unconventional technique to create animations. In these installations, the movement of drawings
painted by invisible ink on transparent acrylics becomes visible under flashes of blacklight LEDs
as animations. Through these exploratory works I seek to find and establish new links of
communications in ways that connect to individuals, including myself, that can create relief and
offer new promises of hope in non-traditional ways
Novel Strategies to Improve the Effectiveness of Field of View - Aware Edge Caching for Adaptive 360° Video Streaming
No full textVirtual Reality (VR) and 360° Video Streaming have attained a lot of popularity recently.
Streaming 360° video to Head Mounted Displays (HMDs) over the internet is extremely
demanding owing to its huge size, desirability to be viewed at higher resolutions, high bandwidth,
and low latency requirements. However, viewers can view only a small portion of a scene in the
video at a time, since viewers are limited by the Field of View (FoV) of the HMD. A few solutions
use adaptive 360° video streaming by streaming high resolution video of only the part on the video
in the viewers FoV, and low-resolution video for the part of the video that is not in the viewers
FoV. FoV Adaptive 360° video streaming has been instrumental in decreasing the bandwidth
requirements, but network latency is another factor that adversely affects the streaming of 360°
videos from distant content servers. To overcome this, some solutions use caching of popular
content at the mobile edge cloud server close to the end user. This caching policy helps reduce
latency in the network and alleviate network bandwidth demands by decreasing the number of
future requests that must be sent to the content server, thus reducing the load on the server. But
most of these strategies use generic heat maps to determine popular content in videos among users.
A viewers’ FoV is a depiction of that viewers’ area of interest in the video at any point in time -
with the center of the FoV being of utmost importance grabbing the viewers’ attention and the
peripheries of the FoV of relatively lesser importance, importance decrease as we move from the
center to the periphery in the FoV. Anything outside the FoV of the viewer is of no importance
since the user chose not to see that part in the video. The importance of a part in the video portrays
its popularity among users – the more the importance, the more the popularity. The popularity of
different parts of the video based on the past viewers’ viewing history determines how significant
each part of the video is to be cached at the edge servers in the above-mentioned caching policy.
More the popularity, the more the probability of the content to be cached. However, the use of
traditional heat maps to determine the popularity of video content gives equal importance to the
entire FoV and fails to cater to the requirement of declining importance given to different parts of
the FoV as we move farther away from the center of the FoV. In this thesis, we show how the use
of such heatmaps gives wrong impression of the popularity of video contents -contents that are not
so popular appear to be popular. This false notion created by heat maps renders them useful only
in highly constricted cases. We demonstrate the relevance of some indices used in election analysis
to overcome this limitation in heat maps and discuss where they would work best. We also
introduce the concept of Vote decay in the popularity of contents in the FoV to remove
misinterpretations of content importance so that we can improve caching decisions and future FoV
predictions
Efficient Nonparametric Spectral Density Estimation with Randomly Censored Time Series
No full textSpectral Density estimation is a well known problem for a directly observed time series.
However, the literature on spectral density estimation for a randomly censored time series is
next to none. The dissertation develops a sharp lower bound for the minimax mean integrated
squared error (MISE) for an estimator of a spectral density for a zero-mean stationary
time series. Then an efficient and data-driven spectral density estimator (E-estimator) is
suggested, which adapts to unknown smoothness of the spectral density and distribution of
a censored random variable. Asymptotic upper bound of the MISE of the proposed estimator
is obtained and it attains the sharp lower bound, so the proposed estimator is sharp minimax
(or efficient). The E-estimator is studied and compared with an Oracle and a Naive estimator
via simulated and real examples. The studies exhibit this E-estimator performs well under
various scenarios and can compete with the Oracle estimator in simulated and real examples
Desistance From Offending: an Examination of the Potential Influences of Adult Institutions
No full textPersistent offending over the life-course is marked by young adulthood as it is this age range that
lies between the two general natural declines of criminal behavior (late adolescence and late
adulthood). Debate within the desistance literature is ongoing into what components are essential
to the initiation and termination of anti-social behavior. One approach argues for turning points
that encourage prosocial conformity while others advocate that a prosocial cognitive shift is
essential. The current study extends the literature by examining the impact of adult institutions
on the criminal desistance process while employing a cognitive element. Findings indicate some
support for the unique contribution of turning points, however, it remains unclear the full impact
of these institutions and identity reformation and where they lie causally within the desistance
process. The implications of these findings regarding policy approaches as well as
recommendations and paths for future research are discussed
Site-Specific PM2.5 Estimation at Three Urban Scales
No full textFine particulate matter, also known as PM2.5, is one of the major risk factors to human health.
Because of their small size, these particles travel deep within human lungs and pose a variety of
health problems. A primary source of acquiring PM2.5 exposure is based on the nearest groundlevel air quality monitoring station. However, these stations are often few and sparsely located
due to their high costs for installation and maintenance. This study addresses three challenges
related to PM2.5. First, the number of air-quality monitoring sites is insufficient to acquire the
complex spatial variability of PM2.5. Therefore, in-situ ground observations fail to characterize
PM2.5 distribution, and hence exposure, adequately. The shortfall calls for models capable of
estimating PM2.5 at unmonitored locations. Satellite-based Aerosol Optical Depth (AOD) serves
as a proxy to estimate PM2.5. Second, although satellite data can supplement PM2.5 estimates at
unmonitored locations, the spatial resolutions of satellite-based estimates of PM2.5 are in the
order of kilometers. These spatial grains are too coarse to capture PM2.5’s spatial variation
caused by contextual geographic factors such as buildings, and subsequently the estimates’
applicabilities to support environmental exposome on health effects. Third, the current standards
measure PM2.5 in terms of mass per volume, but findings from some recent studies suggest that
alternative measures of PM2.5 are also strongly associated with adverse health outcomes.
However, observations in terms of these measures are not available.
The dissertation research aimed to address the three challenges in three studies. The first study
evaluated the potential of the Convolutional Neural Network (CNN) approach to downscale
PM2.5 using satellite-based AOD and meteorological data using Dallas-Fort Worth as a case
study. The study developed a model capable of estimating PM2.5 corresponding to the hour of
satellite overpass time and examined environmental predictors commonly available for all
monitored or non-monitored locations. In particular, the study investigated the effect of the
spatial extent to which predictors from the surrounding area influenced the PM2.5 estimates at a
location. The results showed that the proposed CNN model effectively estimates PM2.5
concentration with correlation coefficient (R) of 0.87 and root mean squared error (RMSE) of
2.57 μg/m3
. Moreover, spatially lagged variables from a wider area around an estimation location
improved the model performance. As most monitoring stations were in open areas, data from
these stations could not be used to examine the effect of contextual factors, such as the building
on PM2.5. The second study evaluated the effects of contextual geographic factors on PM2.5 in
mass per volume (i.e., standard measures) in pedestrian-friendly areas on the University of
Texas at Dallas campus. The study used a mobile sensor to collect spatial and temporal fineresolution PM2.5 data on the campus. The study found very low spatial variation in the study
area less than 1km2
. Furthermore, weather-related variables played a dominant role in PM2.5
distribution as temporal variation over-powered spatial variation in PM2.5 data. The study
employed a fixed effect model to assess the effect of time-invariant building morphological
characteristics on PM2.5 and found that building’s morphological characteristics explained
33.22% variation in the fixed effects in the model. Furthermore, openness in the direction of
wind elevated the PM2.5 concentration. The third study investigated the potential of AOD to
downscale Particle Number (PN) concentration, an alternative measure of PM2.5, and the effect
of building morphology on PN concentration using PN measurements collected across the streets
of San Francisco by the Google streetcar. The study showed that AOD remained useful to
estimate street-level PN concentration across five different particle sizes. The subsequent
analysis of variable importance revealed that AOD and AOD-related variables were more
important than building morphology but less important than meteorological variables in the
estimation of PN concentration
Essays on Production-based Asset Pricing
No full textThis dissertation consists of two essays on production-based asset pricing.
The first essay studies the asset pricing implications of investment and disinvestment op-
tions with a production-based model featuring costly reversibility. Investment options are
contingent claims on assets in place so that they are riskier and earn higher expected re-
turns. Disinvestment options with costly reversibility reduce exposure to aggregate risks
amid deteriorating business conditions and lower expected returns on a firm. The inextri-
cable link between investment options and disinvestment options explains the coexistence of
the profitability premium and the value premium while retains a positive relation between
profitability and market valuation ratios. My model also generates a procyclical profitability
premium and a countercyclical value premium.
In the second essay, my co-authors and I investigate the joint asset pricing effects of variable
costs and fixed costs in a firm’s production process. While the latter such as SG&A expenses
create an operating leverage effect, the variable costs allow firms to hedge against aggregate
profitability shocks. Taking into account both types of production costs explains the empir-
ical patterns in the cross-section asset returns in portfolios sorted by the gross profitability
and operating leverage. Our model reconciles the seemingly contradictory phenomena that
higher productivity firms earn lower returns ( ̇Imrohoro ̆glu and T ̈uzel (2014)), whereas more
profitable, often more productive, firms earn higher returns (Novy-Marx (2013)). It also of-
fers a novel explanation for the negative idiosyncratic volatility premium (Ang et al. (2006))
based on production costs
A Study on Gallium Coating and Wireless Platform for Implantable Biomedical Applications
No full textRecent research interests in wearable or implantable devices have played a significant role in
advancing MEMS technologies into emerging biomedical fields. Novel materials and methods
have been extensively explored in creating intrinsically flexible biomedical devices.
As a novel nontoxic alternative to mercury, gallium-based liquid metals have been utilized to
form functional wearable devices thanks to their unique combination of electrical and fluid
properties. However, the adherent tendency of oxidized liquid metals has been a fundamental
challenge that needs to be addressed in order to unleash the full potentials. This work reports
gallium coating as a simple remedy to convert various microfluidic materials to nonwetting
surfaces against gallium-based liquid metals. Quantitative studies on the super-lyophobicity and
surface topography are presented to evaluate gallium coated surfaces as nonwetting microfluidic
platform for oxidized gallium-based liquid metal droplet manipulation.
Implantable functional devices, on the other hand, require wireless operation capability in order
to reduce invasiveness to biological bodies while fulfilling monitoring or therapeutic functions.
Intramedullary fluid modulation has been reported to enhance bone density and can be employed
as a potential treatment to osteoporosis. Aiming at replacing invasive methodology used in these
in vivo studies, this work presents an implantable and wirelessly operated intramedullary fluid
modulator for on-demand intramedullary fluid modulation. The pressure modulation is evaluated
by theoretical model as well as ex vivo and in vivo experiments.
Additionally, a wireless pressure sensing system with long range transmission capability is
demonstrated as a complement to wireless intramedullary fluid modulator. Details on the system
design is discussed, and evaluation results in terms of pressure response and transmission range
is presented
Essays in Operations Management
No full textThis dissertation consists of three main chapters focusing on operational problems in supply
chain contracting, online food-ordering services, and agricultural open burning in developing
countries.
In Chapter 2, we analyze a contract in which a supplier, who is exposed to disruption risk,
offers a supply-flexibility contract comprising of a wholesale price and a minimum-delivery
fraction ("flexibility" fraction) to a buyer facing random demand. The supplier is allowed to
deviate below the order quantity by at most the flexibility fraction. The supplier's regular
production is subject to random disruption but she has access to a reliable expedited supply
source at a higher marginal cost.
We derive the supplier-led optimal contract and show that supply-chain effciency improves
relative to the price-only contract. More interestingly, even though the buyer lets the supplier
decide how the two share supply risk, profits of both the players increase by the introduction
of flexibility into the contract. Further, supply-flexibility may be even more valuable for
the buyer compared to the supplier. Interestingly, the flexibility fraction is not monotone in
supplier reliability and a more reliable supplier may even prefer to transfer more risk to the
buyer. The robustness of these findings is established on two extensions: one where we study
a buyer-led contract (i.e., the buyer chooses the flexibility fraction) and the other where the
expedited supply option is available to both the supplier and the buyer.
In Chapter 3, We study the problem of managing queues in online food-ordering services
where customers, who place orders online and pick up at the store, are offered a common
quote time, i.e., the promised pick-up time minus the time the order is placed. The objective
is to minimize the long-run average expected earliness and tardiness cost incurred by the
customers. We introduce the family of static threshold policies for managing such queues. A
static threshold policy is one that starts serving the first customer in the queue as soon as
the server is free and the time remaining until the promised pick-up time of that customer
falls below a fixed threshold. In important technical contributions for establishing the attractiveness of the optimal static threshold policy, we develop three sets of lower bounds on
the optimal cost. The first set of lower bounds exploits structural properties of two special
static threshold policies, while the second set utilizes the idea of a clairvoyant optimal policy
by considering a decision maker who has either full or partial knowledge of the outcomes of
future uncertainties. To obtain our third set of lower bounds, we develop bounds on the optimal earliness and tardiness costs by establishing lower and upper bounds on the steady-state
waiting time under an optimal policy. The optimal static threshold policy is asymptotically
optimal in several cases, including the heavy traffic and the light traffic regimes. We also
develop a dynamic threshold policy in which the threshold depends on the queue length. Finally, through a comprehensive numerical study, we demonstrate the excellent performance
of both the static and the dynamic threshold policies.
In Chapter 4, we study how the government can use information-disclosure policies to minimize agricultural open burning in developing countries. Agricultural open burning, i.e., the
practice of burning crop residue in harvested fields to prepare land for sowing a new crop,
is well-recognized as a significant contributor to CO2 and black carbon emissions, and longterm climate change. Low-soil-tillage practices using a specific agricultural machine called
Happy Seeder, which can sow the new seed without removing the previous crop residue, have
emerged as the most effective and profitable alternative to open burning. However, given the
limited number of Happy Seeders that the government can supply, and the fact that farmers
incur a significant yield loss if they delay sowing the new crop, farmers are often unwilling
to wait to be processed by the Happy Seeder and, instead, decide to burn their farms. A
Happy Seeder is assigned to process a group of farms in an arbitrary order. The government
knows, but does not necessarily disclose, the schedule for the Happy Seeder at the start of
the sowing season. Farmers are impatient, in the sense that they incur a disutility per unit
of time associated with waiting for the Happy Seeder. If the Happy Seeder processes a farm,
then the farmer gains a positive utility. At the beginning of each period, each farmer decides
whether to burn her farm or to wait, given the information provided by the government
about the Happy Seeder's schedule. We propose a class of information-disclosure policies,
which we refer to as threshold policies, that provide no information to the farmers about
the schedule until a pre-specified period and then reveal the entire schedule. By obtaining
the unique symmetric Markov perfect equilibrium under any threshold policy, we show that
the use of an optimal threshold policy can significantly lower the number of farms burnt
compared to that under the full-information and no-information disclosure policies
Purification of Cytosolic Sulfotransferases and Towards the Synthesis of an Affinity-based Protein Profiling Tool for 3’- Phosphoadenosine-5’-phosphate
No full textSulfotransferases play a large role in phase II metabolism, which involves the modification of
xenobiotics and endogenous compounds for subsequent elimination from the cell. This is done
through the addition of a sulfuryl group, from the universal sulfate donor 3’-phosphoadenosine-
5’-phosphosulfate (PAPS), replacing the hydroxyl or amine group of a substrate, generating the
sulfurylated product and 3’-phosphoadenosine-5’-phosphate (PAP). Subsequently, the
sulfurylated product can then be exported out of the cell by the associated transmembrane
multidrug resistance transporter (MRP). A mechanistic understanding of these enzymes
mechanism of actions is of vital importance, as disease states, such as anemia, cancer, and autism,
have been linked to dysregulation of sulfurylation. Currently, the state of the art for studying
sulfotransferase activity includes radiometric, fluorometric, photometric, and mass spectrometric
assay methods. The study of sulfotransferases through a chemical approach is comparatively
understudied. This body of work lays the foundation for the study of the activity of these enzymes
through the development of a chemical tool to study its protein interaction partners.
In Chapter 1, a protocol for the expression and purification of six human cytosolic sulfotransferases
is established. In later studies, these enzymes will be utilized to characterize fluorescent
sulfotransferase sensors in vitro prior to applications in live cells. The interaction dependent
fluorescence of the sulfurylated product will allow for the study of sulfotransferase activity. In
Chapter 2, we demonstrate the initial synthetic steps of ((2R,3S,4R,5R)-5-(6-amino-2-(prop-2-yn-
1-ylamino)-9H-purin-9-yl)-4-hydroxy-3-(phosphonatooxy)tetrahydrofuran-2-yl)methyl
phosphate (PAP-A). After treating cells with PAP-A, biotin azide will react with the tag on PAPA,
allowing for isolation through Western blot analysis. Proteins in the isolated product can be
identified through peptide mass fingerprinting. This tool will further our understanding of the role
of PAP by identifying proteins that interact with PAP. Sulfurylation is a vital cellular process and
by studying the enzymes involved, and the by-products of this process, we will improve our
understanding of the roles these components play within living systems