Paul Thielen
I have a strong foundation in Data Science, with a focus on statistics and machine learning. Previously worked as a Data Analyst Intern at Uber, focusing on analytical work and gaining practical experience in data-driven decision-making.
View My LinkedIn Profile
Selected Projects in Data Science, Machine Learning, and NLP
Deep Learning Under Attack: Revealing Vulnerabilities through Shadow Reconstructions - Thesis
My thesis developed a novel hybrid architecture for Model Inversion (MI) (reconstructing inputs to a target CNN via the model outputs) in a gray-box setting (where the target model’s architecture is known but its weights and training data are unknown). It is the first to combine two existing methods: input reconstruction via gradient-based optimization and inverse modeling. Since the target model’s weights and training data are unavailable, a Shadow Model (SM) was trained on a dataset with a similar distribution to approximate the target model’s behavior. Once the SM is trained, its gradient-reconstructed input is combined with its output vector to train an inverse model (a transposed convolutional neural network) using the same training data for the SM. At inference time, the SM approximates the inputs of the target model using the latter's outputs, and the inverse model receives both as input to create a synthetic reconstruction. This hybrid architecture outperformed the gradient-based optimization method on the test set by 54.73%.
MEDUSA: Medical Data Under Shadow Attacks via Hybrid Model Inversion
Accepted to AISTATS 2025 Conference
In collaboration with with the AIMI Center at Stanford I extended upon my thesis by proposing MEDUSA, a novel hybrid model inversion framework leveraging gradient-based optimization and inverse modeling to reconstruct high-fidelity medical images from model outputs in a gray-box setting. We applied the hybrid MI model to reconstruct the MedMNIST datasets and achieved up to 12% improvement in performance compared to standard gradient reconstructions. We demonstrated that models trained on 50% real data supplemented with synthetic reconstructions from our hybrid model performed within 5% of benchmarks set by models trained on 100% real data. This validates the effectiveness of synthetic reconstructions, enabling adversaries to closely replicate clinical models and reconstruct inputs from outputs. We therefore investigated defense mechanisms and found that limiting the publicly available data (e.g. 5% of the total dataset) can significantly degrade (e.g. by 78%) an attacker’s ability to reconstruct the inputs and still maintain data transparency. We also invented a smearing defense technique, whereby a clinical model’s output is reconstructed as a weighted sum of its k-nearest neighbors, reducing reconstruction quality by up to 64%.
South Sudan Food Insecurity
Food insecurity in South Sudan is intensified by conflict, necessitating data-driven solutions. This project, conducted in collaboration with the Zero Hunger Lab, aimed to predict food insecurity risks using local and global news articles about South Sudan and climate data from the World Bank Group. Our group trained linear regression models for two districts, incorporating climate features and the number of articles mentioning violence in the specific district (extracted via GPT-3) to forecast IPC scores monthly. Following the Bentiu takeover, a major conflict event, the models exhibited a 1000% increase in mean-squared error. A Bayesian Structural Time Series model further confirmed a statistically significant 70.82% rise in IPC scores post-takeover, highlighting the lasting impact of conflict on food insecurity. The project earned a 9/10 grade and recognition from the Zero Hunger Lab.
Speech Emotion Decoder
Accurately detecting emotional valence in speech is critical for improving human-computer interactions. This project aimed to develop a model that predicts emotional valence on a 1-5 scale based on audio recordings. Built a 1D CNN regression model in PyTorch, achieving a mean squared error (MSE) of 0.26 on the test set. Optimized performance through hyperparameter tuning, testing optimizers (Adam, Adagrad, SGD with Nesterov momentum), and conducting a random grid search to fine-tune learning rates and hidden layer configurations.
Airline Customer Service Analysis - Twitter
Effective social media engagement is critical for maintaining customer satisfaction and loyalty in the airline industry. This project aimed to evaluate KLM’s Twitter-based customer service and determine if it provides a competitive advantage. Analyzed 6.5M tweets using a fine-tuned BERT model for sentiment analysis and RAKE for topic extraction, and developed a KPI-based ranking system incorporating sentiment trends, response time, and case resolution rates. We as a team brought forth the innovative idea of using the proportions of positive/negative tweets around key topics as metrics for competitive analysis.
Predicting Burglaries in Barnet, UK
The Barnet borough of London faces challenges in allocating limited resources to combat burglaries effectively. This project aimed to forecast burglary trends and optimize police resource allocation. Implemented OLS, Random Forest, SARIMA, and Prophet models, training a model for each Lower Super Output Area (LSOA) to account for regional differences. The Prophet model demonstrated the highest accuracy, achieving an average R^2 score of 0.83, effectively capturing 83% of the variability in burglary rates across the regions. Additionally, used the Gurobi Integer Linear Programming (ILP) solver to optimize officer allocation across LSOAs. The project provided data-driven strategies to improve police officer deployment in Barnet to reduce burglary rates.
AirBnB Dashboard
As Airbnb’s popularity grows, understanding complex markets like New York is essential for property investors and hosts. This group project aimed to create an interactive Plotly dashboard to provide the mentioned stakeholders with a competitive edge. The dashboard featured linked visualizations, including a choropleth map of New York, bar charts, box plots, scatter plots, and a Parallel Coordinates Plot, allowing users to explore connections between various aspects of the market. Spatial data was integrated to calculate mean distances to transit and recreational activities within one mile of each listing. The dashboard allows stakeholders to make data-driven decisions on pricing, policies, and market positioning.
Spread of Misinformation
An experiment by Aslett et al. 2022 found that approximately one-third of the participants visited unreliable news websites. Misinformation on Social Media is a serious problem that can have varying harmful impacts depending on the severity of the false information. This project aimed to simulate the spread of misinformation by modeling user interactions and peer-to-peer news sharing. My group and I developed a multi-agent simulation in NetLogo and implemented a generalized least squares regression to analyze the influence of source reliability and news severity on the compromised population percentage. Experiments revealed that higher source reliability increased the compromised population rate by up to 37%, while the severity of the news reduced the time to reach 95% of the population by 40%. These findings highlight the significant influence source credibility and news severity have on the rapid spread of misinformation.
