A paper about privacy (or lack thereof) in Nature by Knolle et al. about medical AI models that exhibit a weakness to membership privacy attacks thru multiple queries of the public model. As discussed in a commentary article by Zhang & Ghassemi, a membership privacy attack proves successful when the confidence of the model prediction jumps to higher values for a real target of interest compared with an imaginary one. This obviously assume that users (and attackers) may repeat queries ad nauseam from the model, which differs from our Bayesian privacy setting where the output is provided once and only once (whatever the release mechanism is).
The attacker is modelled as resorting to a basic likelihood-ratio MIAs3,4, that is, a test based on the prediction confidence attached to the target model with the null being that the target is not a member:
“the parameters of the distributions under the two hypotheses are specified by parametric fitting of sample confidence values obtained from reference models. Reference models are models assumed to be trained by the attacker and are ideally, but not necessarily, of similar architecture as the target model and trained on data similar to the training dataset” – M Knolle & al.
but the paper does not provide further details about inferring about the model parameters. The following sentence is also unclear
“objectively larger threats are posed by privacy attacks with stronger assumptions on a potential attacker, such as access to model parameters17, access to parameter updates during model training18 (…) By contrast, the type of attack we consider here requires querying the target model only once (to obtain a prediction for the target record)” – M Knolle & al.
in that, indeed, returning the model parameter estimates is more informative about the data than a black-box prediction interface, but I miss the single query point as it seems to me that the attacker must multiply the queries to build a confidence distribution under both hypotheses.
“Purely technical measures, such as a mathematical approach called differential privacy9, often create performance trade-offs10 that are too limiting for medical AI tools. Instead, what is needed are regulatory and sociotechnical safeguards, such as privacy audits and risk assessments, that are specific to the domain in question.” – H Zhang & M Ghassemi
“our results indicate that privacy attacks against AI models may be much more effective at compromising the privacy of individual data contributors than previously thought. This suggests that current AI privacy risk reporting practices may underestimate individual-level risk and thus motivates the integration of mathematically verifiable risk mitigation strategies such as differential privacy (DP) into medical AI model development workflows.” – M Knolle & al.
“the finding that record-level differential privacy is insufficient for multi-record patients is particularly impactful and has clear policy and implementation implications” –Referee #2
“we also found that the full mitigation of MIAs for all data-contributing patients requires stricter levels of privacy protection (ϵ, δ ) smaller than previously believed. Moreover, our results also show that fully mitigating MIAs requires DP accounting at the patient- rather than the record-level.” – M Knolle & al.
In a funny (?) clash, the authors of the paper and those of the comment and a referee seem to disagree on the pertinence of differential privacy guarantees in this context. But the conclusion remains very vague on a rigorous way to assert privacy leaks and confidentiality protection for a given AI and its supporting dataset.
