AI 용어집
인공지능 완전 사전
Epsilon-Differential Privacy
Quantitative measure of privacy where epsilon (ε) represents the privacy loss parameter, with lower values indicating stronger protection but reduced data utility.
Delta-Differential Privacy
Variant of differential privacy allowing a probability δ of privacy violation, used for mechanisms with weaker privacy guarantees but better utility.
Laplace Mechanism
Basic algorithm for achieving differential privacy by adding noise drawn from a Laplace distribution calibrated according to the query function's sensitivity and epsilon parameter.
Gaussian Mechanism
Alternative to the Laplace mechanism using a normal distribution to add noise, particularly suitable for vector queries and offering better composition properties.
Sensitivity
Measure of the maximum impact a single individual can have on the result of a function, fundamental for calibrating the amount of noise needed in differential privacy mechanisms.
Privacy Budget
Total amount of privacy loss (epsilon) allocated for a series of queries on sensitive data, managed to maintain overall privacy guarantees.
Randomized Response
Historical technique for collecting private data where respondents randomly lie or tell the truth according to a predefined probability, precursor to modern local differential privacy methods.
Composition Theorem
Mathematical principle defining how privacy guarantees compose when multiple differentially private mechanisms are applied sequentially to the same data.
Local Differential Privacy
Privacy model where noise is added directly to the user's data before collection, eliminating the need to trust a centralized data collector.
Global Differential Privacy
Approach where a trusted administrator applies differential privacy mechanisms on the complete database, generally offering better utility guarantees than local privacy.
Private Aggregation
Techniques combining individual noisy contributions to produce privacy-preserving aggregated statistics, essential for differentially private counts and averages.
Noise Calibration
Mathematical process determining the optimal amount of noise to add based on the function's sensitivity and the desired privacy level to maximize the utility of results.
Post-processing Invariance
Fundamental property ensuring that applying any additional function to the result of a differentially private mechanism does not degrade the privacy guarantees.
Exponential Mechanism
General differential privacy mechanism for non-numeric queries, selecting outputs according to an exponential distribution weighted by their quality with respect to the data.
Rényi Differential Privacy
Generalization of differential privacy using Rényi divergence to measure privacy loss, offering more precise composition analyses and better intermediate guarantees.
Differential Privacy Amplification
Phenomenon where privacy guarantees naturally improve when mechanisms are applied to random subsamples of data or combined with stochastic processes.
Sparse Vector Technique
A differentially private method that allows determining whether a set of thresholds is exceeded without revealing the exact values, useful for adaptive analyses and pattern discovery.