LiRA — Membership Inference from First Principles (Carlini et al.)
Why
Defines the evaluation standard the thesis should eventually meet. The simple loss-threshold MIA in
the ER notebook is a population attack; LiRA explains why that is not enough and how to do better.
Summary
Bibliographic. Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis,
Florian Tramèr (Google Research / UMass Amherst). Membership Inference Attacks From First
Principles. arXiv:2112.03570v2, 12 Apr 2022. 20 pages.
The critique
Membership inference is usually evaluated with average-case accuracy on a balanced set of
members and non-members. The paper argues this is the wrong metric:
- Accuracy is symmetric: it weights false positives and false negatives equally, but a privacy
attacker cares about confidently identifying a few members with very few false positives. - An attack that identifies 0.1% of members with no false positives can have the same average
accuracy as an attack that is barely better than random. - The right metric is TPR at a fixed low FPR (e.g. 0.1% or 0.001%). Many prior attacks have
0% TPR at 0.1% FPR; “average accuracy” hid this.
Method
LiRA casts membership inference as a likelihood-ratio test:
- Train many shadow models on random subsets of the data, so that each example is sometimes a
member (IN) and sometimes not (OUT). - For a target example , collect the model’s confidence in each shadow model
and fit two distributions, and . - Apply a logit scaling to the confidence: only the logit is well approximated by a pair of
Gaussians (raw confidence and cross-entropy loss are not). - Compute the likelihood ratio of the target model’s confidence under vs and
threshold it. The offline variant trains shadow models on random datasets ahead of time,
never on the target point; it estimates only and uses a one-sided
hypothesis test. The online variant trains new shadow models for the query and fits both the
IN and OUT distributions. - Parametric Gaussian modelling needs far fewer shadow models than nonparametric estimation — the
paper matches prior work with 400× fewer models.
Results
- 10× more powerful at low FPR than prior attacks, and strictly dominates them on existing
metrics. - On a non-overfit CIFAR-10 ResNet (92% test accuracy), LiRA reaches 63.7% balanced accuracy versus
60.0% for the best prior attack (Ye et al.); the gap is much larger at low FPR. - The attack exposes per-example hardness: some examples are outliers that a model memorises
strongly even when the model is not overfit overall.
Limitations
- Shadow models must match the target training distribution and algorithm; for large models this is
expensive. - The attack assumes the adversary can train many models on similar data.
- The paper’s experiments are image classifiers; generative models are not covered.
Thesis hook
- The ER notebook’s MIA (per-node NLL, threshold, AUC) is a population attack. Per the
inexact-unlearning paper, it will overestimate forgetting. The thesis should either upgrade to a
LiRA-style per-example attack or explicitly label its MIA as weak evidence. - LiRA gives the correct vocabulary for the evaluation section: TPR@FPR, shadow models,
per-example likelihood ratio. Reviewers will expect this if the thesis claims privacy. - For graphs, the natural adaptation is per-graph or per-node confidence under shadow models; the
shadow-model cost is the main obstacle. - Per-example hardness connects to the topology question: are hub nodes / high-degree graphs
systematically easier to identify (and harder to forget)?
Open questions
- How many shadow models are needed for a graph diffusion model, and can a smaller GNN surrogate
stand in? - What is the graph analogue of “confidence”: likelihood of the graph, edge-level log-likelihood,
or a kernel-based score? - Does LiRA’s per-example hardness correlate with degree, clustering, or graph family? That would
turn the evaluation into a topology result.