Muhammad Usama Saleem

Muhammad Usama Saleem

Graduate Student

The University of North Carolina at Charlotte

About me

I am a Graduate student at the University of North Carolina at Charlotte doing research in the Data Privacy and Deep Learning.

Interests

  • Software Engineering
  • Data Science
  • Machine Learning
  • Security
  • Data Privacy

Education

  • Computer Science, 2021-Present

    The University of North Carolina at Charlotte, U.S

  • B.S in Computer Science, 20127-2021

    Lahore University of Management Science (LUMS)

Skills

python

C

cplusplus

csharp

java

matlab

scala

mysql

bash

latex

R

haskell

docker

git

jupyter

numpy

opencv

pandas

pytorch

scikit

tensorflow

huggingface

Recent Publications

Private Data Synthesis from Decentralized Non-IID Data

Under Review in Artificial Intelligence and Statistics (AISTATS), Spring 2023, Valencia, Spain
Muhammad Usama Saleem , L. Fan
Private Data Synthesis from Decentralized Non-IID Data

Privacy Enhancement for Cloud-Based Few-Shot Learning

Accepted in IEEE WCCI International Joint Conference on Neural Networks (IJCNN), July 2022, Padua, Italy
A. Parnami, Muhammad Usama Saleem , L. Fan, M. Lee
PDF
Privacy Enhancement for Cloud-Based Few-Shot Learning

DP-Shield: Face Obfuscation with Differential Privacy

Accepted in Proceedings 25th International Conference on Extending Database Technology (EDBT 2022) Edinburgh, UK, March 29 - April 1, vol. 25, pp. 578–581, Mar. 2022.
Muhammad Usama Saleem , D. Reilly, L Fan
PDF Code
DP-Shield: Face Obfuscation with Differential Privacy

Work Experience

 
 
 
 
 

Graduate Research Assistant

Data Privacy Lab

Sept. 2021 – Present Charlotte, NC
  • Developed solutions for privacy-preserving data sharing & reduced privacy risks by 91% in SOTA deep learning systems
  • Achieved a 5x decrease in computational efficiency by introducing federated learning for GANs model for multi-clients
  • Reduced the re-identification attack rate by 85% of adversary possessed facial recognition system and achieved 75% utility by designing & implementing image obfuscation method for social media data
 
 
 
 
 

Data Scientist

Afiniti

June 2021 – Aug 2021 Lahore
  • Increased sales of clients by 6% by optimizing AI models to determine how routing of call agents & customers will be conducted in call centers. Moreover, we scrutinized the performance of AI models in real-time to maximize productivity in terms of sales.
Read more about Afiniti. here.
 
 
 
 
 

Software Engineer Intern

Technos

May. 2021 – Aug. 2021 Lahore
  • Increased website response time by 10% by redesigning REST-API & minimizing HTTP requests for 1M+p.a. traffic platform
 
 
 
 
 

Software Engineer Intern

BrainTechs

Jan. 2020 – Dec. 2020 Lahore
MobileMechanic - Auto-Repair Application
  • Co-founded & developed first MobileMechanic platform in Pakistan that allows customers to communicate with mechanics
  • Collaborated in an agile way with a team of experienced developers on requirement collection, development & unit testing.

Projects

 
 
 
 
 

Fairness & Biases in Facial Recognition & Obfuscation Methods

UNC Charlotte

June 2022 – Present Charlotte, NC
  • Investigated the facial recognition & obfuscation method biases toward particular demographics like race, gender & age
  • Discovered White Males & Females have 13% and 10% more re-identification risks as compared to Asian Males & Females
 
 
 
 
 

Spotizer

UNC Charlotte

Jan. 2022 – May 2022 Charlotte, NC
  • Designed a speech diarization ML model to split audio streams according to speaker identity & reduced its error rate by 3%.
  • Incorporated noise removal, speech-to-text, & language detector for speakers with 97.5% accuracy to assist users
Read more about Afiniti. here.
 
 
 
 
 

Making User’s Check-Ins Secure in Location-Based Social Networks

Privacy Lab

Sept. 2021 – Dec. 2021 Charlotte, NC
  • Designed a cutting-edge model to learn latent representations of user check-ins to generate fake locations for privacy preservation of location-sensitive activities. Moreover, the paper is to be submitted at SIGSPATIAL 23.
  • Outperformed state-of-the-art models by 10% & 12% in terms of precision & recall by providing diversified recommendations
 
 
 
 
 

BingeIT

LUMS

Jan. 2021 – May. 2021 Lahore
  • Developed a movie streaming app & content-based recommender system with 95% precision@50 to make relevant suggestions
 
 
 
 
 

Classification & evasion of encrypted traffic fingerprinting

Internet Measurements Lab

June 2020 – Dec 2020 Lahore
  • Surpassed SOTA models with 97% accuracy by implementing android app identification via encrypted network traffic analysis
  • Reduced accuracy by 55% by designing robust attacks for adversaries to evade app fingerprinting using adversarial ML
 
 
 
 
 

Investigating the compliance with COPPA in browser-based games

Internet Security Lab

Jan. 2020 – May. 2020 Lahore
  • Discovered 55 out of 300 children’s gaming platforms exposed pornographic material to underage users and collected personally identifiable information (PII), which is utilized for targeted advertising.
  • Presented findings to the COPPA body along with actionable insights to prevent adult content
 
 
 
 
 

Minimizing Spread of Misinformation by Countering Deepfakes |

Center for Speech and Language Technologies

May 2019 – Dec. 2019 Lahore
  • Investigating dissemination of fake news factors in low digital-literacy people when they are exposed to deep-fake audios
  • Achieved a 4.2 mean opinion score for first Urdu text-to-speech system for generating realistic fake audios of politicians

Teaching Experience

UNC Charlotte

 
 
 
 
 

Teaching Assistant

CS 437 - Deep Learning

Jan. 2021 – May 2021 Dr. Murtaza Taj - Queen Mary University of London
Developed and evaluated assignments, given tutorials, held office hours, and tutored over 130 undergraduate and graduate students.
 
 
 
 
 

Teaching Assistant

CS 535 - Machine Learning

Sept 2020 – Dec 2020 Dr. Agha Ali Raza - Carnegie Mellon University
Designed and graded projects conducted tutorials, and held office hours for over 120+ undergrads, and graduates