Differentially Private Algorithms For an algorithm to be differentially private, its output should not change even if a data point is excluded from the dataset. This provides confidence that even if personally identifiable information is present within the dataset, it would not be visible to the outside world. DP algorithms are resistant to adaptive attacks since the noise introduced into the dataset makes the data imprecise.
DP Algorithms Models and Explanation
How is a model made differentially private?
MODELS
Advantages
Disdvantages
Slow at training. Overfitting. Not suitable for small samples. Small changes in training data change the model. Occasionally too simple for very complex problems.
High accuracy. Good starting point to solve the problem. Flexible and suitable for a variety of different data. Fast to execute. Easy to use. Can model missing values. High performing.
Exponential Mechanism adds noise to the prediction probability of labels regarding the most frequent label.
Tree based algorithms
Easy to learn, configure and maintain. Simple to implement.
Inconsistent (depends on the selection of the initial seed). The âKâ input requires specifying the size of the clusters.
Noise is added to the averages of centroids calculated where noise is taken from a Laplace distribution, which is a function
Unsupervised Learnings
Aims toward spherical clusters (for some applications might be a con). Handles large datasets.
of the number of centroids, epsilon, sensitivity, and the number of data partitions.
Sensitive for outliers, especially if they were used as initial seeds
Easy to implement, the theory is simple, low computational power compared to other algorithms. Easy to interpret coefficients for analysis. Perfect for linearly separable datasets. Inclined to overfit, but can be avoided using dimensionality reduction, cross-validation, and regularization techniques.
Laplacian noise is added to the coefficients of the objective function. Noise is added to the coefficients of each feature where noise is proportional to the exponential function.
Prone to underfitting.
Sensitive to outliers.
Linear Models
Assumes that the data is independent.
5
© 2023 Fractal Analytics Inc. All rights reserved
Made with FlippingBook - PDF hosting