ML Engineer in progress · Privacy-aware deep learning · Open source contributor
github.com/pragati-0208/dp-ml-privacy
Implemented DP-SGD training on CIFAR-10 using PyTorch + Opacus. Measured the privacy-accuracy tradeoff across epsilon values and demonstrated membership inference attack resistance — showing that DP training reduces attacker advantage from AUC ~0.73 to near-random.
| Metric | Value |
|---|---|
| Baseline accuracy (no DP) | 74.29% |
| DP accuracy (ε=0.5, σ=2.46) | 39.93% |
| DP accuracy (ε=1.0, σ=1.41) | ~54% |
| Attack AUC without DP | ~0.73 |
| Attack AUC with DP | ~0.54 |
Stack: Python · PyTorch · Opacus · Scikit-learn · Matplotlib
Python PyTorch Opacus Pandas NumPy Scikit-learn Matplotlib Seaborn EDA Feature Engineering Git HTML CSS
GSSoC Contributor
INCF / knowledge-space-agent
- Replaced
print()calls with structured Python logging module · PR #75 · 4 comments received - Fixed AgentState dunder keys · PR #84 · merged
- Added backend test suite and pytest configuration · PR #82
NeurodataWithoutBorders / nwbwidgets
- Fixed pandas FutureWarning in
infer_categorical_columns· PR #323 - Eliminated test warnings across the codebase
INCF / csa
-
Added Windows & Anaconda installation docs · PR #27
-
Made project accessible to Windows users
Open to internships · ML roles · open source collaborations