Alibi Detect

Beyond detection tasks, alibi_detect provides several model implementations in alibi_detect.models that can be used as building blocks. The main TensorFlow-based implementations include PixelCNN, VAE (Variational Autoencoder), Seq2Seq (Sequence-to-sequence), and resnet.

from alibi_detect.models.tensorflow import PixelCNN
from alibi_detect.models.tensorflow import VAE
from alibi_detect.models.tensorflow import Seq2Seq
from alibi_detect.models.tensorflow import resnet

Alibi-detect is integrated with the following platforms:

• Seldon Core: Supports outlier and drift detection.
• KFServing: Supports outlier and drift detection.

The alibi_detect.datasets module provides utility functions to fetch various datasets for different modalities (Sequential, Images, Tabular). Depending on the function call, you can receive either the data and labels directly or a Bunch object containing data, labels, and optional metadata. Use the return_X_y=True parameter to get a tuple of (X, y) or (X_train, y_train), (X_test, y_test).

from alibi_detect.datasets import fetch_ecg

(X_train, y_train), (X_test, y_test) = fetch_ecg(return_X_y=True)

The alibi-detect API is organized into several functional modules. You can find detailed documentation for each module in the API Reference. Key modules include:

• alibi_detect.ad: Adversarial detection algorithms (e.g., adversarialae, model_distillation).
• alibi_detect.cd: Concept Drift detection algorithms, including base classes, online detectors, and framework-specific implementations (PyTorch, TensorFlow, Scikit-learn).
• alibi_detect.od: Outlier Detection algorithms (e.g., ae, isolationforest, vae, mahalanobis).
• alibi_detect.models: Model architectures and trainers for PyTorch and TensorFlow.
• alibi_detect.saving: Utilities for saving, loading, and validating detectors and their schemas.
• alibi_detect.utils: A wide range of utility modules for data handling, distance metrics, framework-specific helpers, and visualization.

Model distillation is a technique used to detect adversarial data, malicious data drift, or data corruption by comparing the output distributions (logits) of an original large model with a smaller, distilled model.

In this context:

• Harmful data points: Inputs where the original model is correct on uncorrupted data but incorrect on corrupted data.
• Harmless data points: Inputs where the original model remains correct even on corrupted data.

The detector computes an adversarial score $S(x)$ based on the loss function used for distillation between the original and distilled models. If $S(x)$ exceeds a threshold, the instance is flagged as adversarial.

The Spectral Residual (SR) outlier detector is designed for unsupervised online anomaly detection in univariate time series data.

It works by:

1. Computing the Fourier Transform of the input data.
2. Calculating the spectral residual of the log amplitude of the transformed signal.
3. Applying the Inverse Fourier Transform to create a saliency map in the time domain.
4. Computing an anomaly score based on the relative difference between the saliency map values and their moving averages.

If the score exceeds a specified threshold, the timestep is flagged as an outlier.

The Variational Auto-Encoder (VAE) is an unsupervised or semi-supervised outlier detector suitable for tabular and image data. It is trained on a batch of unlabeled, normal (inlier) data.

Core Mechanism: The detector attempts to reconstruct the input data. If an input cannot be reconstructed well, it results in a high reconstruction error, and the data is flagged as an outlier.

Scoring Methods: Reconstruction error is measured using one of the following:

1. Mean Squared Error (MSE): The MSE between the input and the reconstructed instance.
2. Probability: The probability that both the input and the reconstructed instance are generated by the same process.

Currently, the score_type parameter only supports the default 'mse'.

Alibi Detect provides several algorithms for outlier detection across different data types. Choose based on your data format:

• Tabular: Isolation Forest, Mahalanobis Distance, AE, VAE, AEGMM, VAEGMM, Likelihood Ratios.
• Image: AE, VAE, AEGMM, VAEGMM, Likelihood Ratios.
• Time Series: Likelihood Ratios, Prophet, Spectral Residual, Seq2Seq.
• Text: (None specifically listed in the table, but check reference list).
• Categorical Features: Isolation Forest, Mahalanobis Distance, Likelihood Ratios.

Some detectors support Online detection (e.g., Mahalanobis Distance, Spectral Residual) or Feature Level detection (e.g., AE, VAE, Likelihood Ratios, Spectral Residual, Seq2Seq).

The Likelihood Ratio (LLR) outlier detector uses the log-likelihood ratio between two generative models as an outlier score. One model is trained on the original data, while a second model (model_background) is trained on a perturbed version of the dataset. This approach aims to capture background statistics while erasing semantic features through perturbations (e.g., using a Bernoulli distribution to substitute features).

For images, the package provides a PixelCNN++ implementation available in alibi_detect.models.tensorflow.pixelcnn.