KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Differential Cross-Validation
Model evaluation technique that incorporates differential privacy to ensure results do not reveal any information about individual training data. This approach adds statistically controlled noise to performance metrics to protect privacy while maintaining acceptable evaluation accuracy.
Private Split Conformal Learning
Non-conformist evaluation method that divides data into calibration and test sets, then applies privacy preservation techniques on the calibration set. This approach allows generating valid prediction intervals while protecting original data against disclosure.
Federated Cross-Validation
Distributed evaluation protocol where data remains on its original servers and only the model or its metrics are shared securely between participants. This method enables collaborative cross-validation without centralization of sensitive data.
Private Ensemble Evaluation
Evaluation technique using ensembles of models trained on different private partitions of data to produce privacy-preserving aggregated metrics. The approach combines predictions securely without exposing individual model contributions.
Homomorphic Cryptography Cross-Validation
Advanced method allowing cross-validation operations to be performed on encrypted data without ever decrypting it during the evaluation process. This technique ensures absolute mathematical protection of data while enabling precise calculations on performance metrics.
Cross-Validation with Laplace Noise
Differential privacy preservation approach adding random noise following a Laplace distribution to cross-validation metrics. The noise level is calibrated according to the epsilon parameter to guarantee a balance between privacy and result utility.
Data Masking Cross-Validation
Technique where sensitive attributes are masked or transformed before cross-validation while preserving the statistical properties necessary for evaluation. This method uses generalization or selective suppression techniques to protect record identities.
Secure Multi-Party Cross-Validation
Cryptographic protocol allowing multiple parties to collaborate on cross-validation without revealing their respective data to other participants. Each party contributes to the global computation while maintaining their data confidential through MPC protocols.
Perturbation Cross-Validation
Protection method where data is intentionally perturbed through mathematical transformations before the cross-validation process. The perturbation is designed to preserve statistical distributions while making identification of original records impossible.
Secure Aggregation Cross-Validation
Technique where local performance metrics are calculated on private partitions then securely aggregated without revealing individual contributions. The aggregation uses cryptographic protocols to combine results while preserving the confidentiality of each source.
K-anonymous Cross-Validation
Approach ensuring that each record in validation sets is indistinguishable from at least k-1 other records. This k-anonymization technique is applied before cross-validation to prevent re-identification attacks.
Synthetic Generation Cross-Validation
Method using statistically generated synthetic data from original data to perform cross-validation. The synthetic data preserves essential statistical properties while being mathematically decorrelated from real records.
Private Partitioning Cross-Validation
Technique where data is partitioned according to privacy-preserving schemes before cross-validation, ensuring no partition contains identifiable information. The partitioning is optimized to maintain statistical representativeness while protecting privacy.
Encapsulation Cross-Validation
Method where training and test data are encapsulated in secure execution environments (enclaves) during cross-validation. The encapsulation ensures that even the host operating system cannot access plaintext data during evaluation.
Obfuscation Cross-Validation
Protection technique applying obfuscation transformations to data before cross-validation, making their direct interpretation impossible while preserving their utility for evaluation. The obfuscation is reversible only with appropriate cryptographic keys.
Splitting Cross-Validation
Privacy-preserving approach where sensitive attributes are split between different tables or partitions before cross-validation. This technique ensures that no partition alone contains complete identifiable information.
Cross-Validation by Cryptographic Wrapping
Advanced method where data is wrapped in layers of cryptographic protection before cross-validation, with each layer being able to be removed sequentially for graduated access levels. This hierarchical approach allows for granular control of information disclosure.