Membership inference fails using image frequency detail to detect memorization

Frequency-Domain Membership Inference Does Not Beat Confidence Thresholding: A Band-Resolved Analysis of Why

Artificial Intelligence

Summary

Membership inference tries to guess whether a certain data point was used to train a machine learning model by seeing how the model’s confidence changes when parts of the input are removed. The authors tested if removing high-frequency details in images helps spot memorized samples better. They found this approach does not work better than simply looking at how confident the model is overall. The difference between members and non-members shrinks as image detail is lost, making the method less reliable.

What this means in practice

  • For machine learning engineers: Avoid using frequency-based input filtering for membership inference attacks or defenses, since it performs worse than confidence threshold methods.
  • For privacy protection teams: Focus privacy defenses on overall confidence patterns rather than attempts to exploit frequency band-specific model behavior.

Authors

Mohammad Zare, Pirooz Shmamsinejadbabaki

Abstract

A widely repeated account of membership inference holds that models lean on fine-grained detail for samples they have memorized, so degrading an input's resolution should expose membership through how quickly a prediction falls apart. We test that account band by band. A monotone family of zero-phase low-pass filters is applied to each query image, the target's confidence on its original prediction is read at every step, and the resulting decay is scored in a dozen ways. Two federated ResNet-18 targets are studied, one trained normally and one pushed to a 45.1-point generalization gap. The account does not survive. On the normally trained target the best profile score reaches an AUC of 0.513 against 0.537 for a single confidence threshold, and on the overfit target 0.735 against 0.943. A search over 1206 candidate scores, with direction chosen on a held-out half and a fitted combination as an upper bound, produces no score that beats the threshold in either regime. The failure has a mechanism: the mean confidence gap between members and non-members grows as detail is removed, exactly as the hypothesis predicts, but the spread grows faster, so discriminability falls from the first step onward. Decomposing each image into phase and magnitude shows why no band-localized operation can work, since the advantage travels with image structure, surviving at 0.567 when the phase is kept and collapsing to 0.492 when only the magnitude is. Membership is a whole-image identity effect rather than an accumulation of spectral energy in any band, and defenses aimed at high-frequency reliance are aimed at the wrong target.