
Ali Mohammadi Ruzbahani · Zenodo (CERN European Organization for Nuclear Research) 2026 · 2026
DOI: 10.5281/zenodo.24210998
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Security monitoring needs visibility into user behavior, which is a privacy risk for the people observed. Existing trade-off formulas cover one noise mechanism each and ignore the privacy lost when an attack goes undetected. Detecting an attack and inferring a person's behavior are hypothesis tests on the same monitoring record, so its visibility $\mu$, a Gaussian differential privacy parameter, bounds the power of both. An attack of salience $\kappa$, its telemetry footprint in privacy units, is missed with probability at least $\Phi(z_\alpha-\kappa\mu)$ at false-alarm budget $\alpha$, where $\Phi$ is the normal distribution function and $z_\alpha=\Phi^{-1}(1-\alpha)$. Privacy is lost through the record and through the breaches that monitoring fails to stop. With both channels the security-privacy frontier has a closed form, and the sign of its slope, the exchange rate, separates conflict from complementarity. A pressure ratio and a critical salience classify every regime. Privacy alone calls for monitoring exactly when the ratio exceeds a threshold, which equals one for an attack the size of one privacy unit. Above it the optimal visibility is explicit; custodial trust and breach severity enter it through a logarithm. Against an attacker who chooses its own salience, the worst-case undetected impact is at least an explicit constant times $\mu^{-\gamma}$, where $\gamma$ is the elasticity of impact to salience. Gaussian monitoring attains the bound, and the attacker is then caught with a probability independent of $\mu$. Simulations of Gaussian, Laplace and logistic mechanisms agree with the closed forms to within sampling error.
No comments yet — start the discussion below.