Privacy Testing for Deep Learning

A comprehensive toolbox for model inversion attacks and defenses, which is easy to get started.

[ICML 2022 / ICLR 2024] Source code for our papers "Plug & Play Attacks: Towards Robust and Flexible Model Inversion Attacks" and "Be Careful What You Smooth For".

reveal the vulnerabilities of SplitNN

Code for "Variational Model Inversion Attacks" Wang et al., NeurIPS2021

Research into model inversion on SplitNN

📄 [Talk] OFFZONE 2022 / ODS Data Halloween 2022: Black-box attacks on ML models + with use of open-source tools

My attempt to recreate the attack described in "Model Inversion Attacks that Exploit Confidence Information and Basic Countermeasures" by Fredrikson et al. in 2015 using Tensorflow 2.9.1

Simulates fMRI signals with Dynamic Causal Modeling. Demonstrates model degeneracy through model inversion. Addresses this issue with BENCH (Bayesian EstimatioN of CHange).

Code for VidModEx: Interpretable and Efficient Black Box Model Extraction for High-Dimensional Spaces

Implementation of "An Approximate Memory based Defense against Model Inversion Attacks to Neural Networks" and "MIDAS: Model Inversion Defenses Using an Approximate Memory System"

Experiments at the intersection of ML security & privacy: adversarial attacks/defenses (FGSM/PGD, adversarial training), differential privacy (DP-SGD, ε–δ), federated learning privacy (secure aggregation), and auditing (membership/model inversion). PyTorch notebooks + eval scripts.