tinfoilhat0/memberinference-by-lossthreshold

An implementation of loss thresholding attack to infer membership status as described in paper "Privacy Risk in Machine Learning: Analyzing the Connection to Overfitting" (CSF 18) in PyTorch.

MemberInference-by-LossThreshold

This is a minimal implementation (in Pytorch 1.11) of loss thresholding attack to infer membership status (i.e., part of training data or not) of data samples. Briefly, we have a threshold T (e.g., avg. training loss), and we predict a sample 'member' if sample loss < T.

Attack code is in src/attacks.py and some results on common image datasets (averaged over 5 runs) are as follows. Note that, baseline balanced acc. is 50% for MIA.

Dataset	Model	Test Acc (%)	Generalization Gap (%)	MIA Balanced Acc. (%)
MNIST	LeNet	99.2 ± 0.05	0.76 ± 0.04	50.34% ± 0.1
Fashion-MNIST	LeNet	90.77 ± 0.11	5.02 ± 0.63	51.85 ± 0.16
CIFAR-10	ResNet20	87.25 ± 0.39	6.8 ± 0.8	52.47 ± 0.42

Simply run src/runner.sh to generate logs, and then run notebooks/tensorboard_log_reader.ipynb to average the results of logs.