Frequency-Domain Membership Inference Fails to Outperform Confidence Thresholding

Testing across different frequency bands shows that degrading image resolution does not improve membership inference attacks compared to simple confidence thresholding. On ResNet-18 models, frequency-based scoring achieved lower AUCs than standard thresholding methods for both normally trained and overfit targets.