Out-of-Distribution (OOD) Detection

The foundational approach uses the model's own confidence scores as a proxy for familiarity. The core assumption is that a model will produce a lower maximum softmax probability for inputs far from its training manifold.

• Mechanism: A threshold is set on the predicted class probability; inputs falling below this threshold are flagged as OOD.
• Limitation: Modern deep networks are often overconfident on nonsensical inputs, making raw softmax scores an unreliable baseline without recalibration.

Rather than trusting the final output layer, these methods analyze the geometric structure of the model's penultimate layer embeddings.

• Mahalanobis Distance: Fits a class-conditional Gaussian distribution to the training embeddings and flags inputs that are far from all class centroids.
• K-Nearest Neighbors (KNN): Computes the distance to the k-th nearest training sample in the feature space; a large distance indicates a novel region.
• Advantage: These methods are non-parametric and do not require retraining the classifier.

An energy function maps an input to a single scalar, the Helmholtz free energy, where in-distribution data is assigned low energy and OOD data high energy.

• Energy Score: Defined as the negative log of the denominator in the softmax function, aligning with the model's logit values.
• Differentiation: This approach creates a sharper separation between ID and OOD samples compared to softmax scores because it leverages the unnormalized logit space, avoiding the saturation issues of probability normalization.

Sophisticated detectors analyze the input itself for statistical anomalies before the model processes it.

• Likelihood Regret: Compares the likelihood of an input under a generic background model versus a model fine-tuned on the specific training distribution.
• Typicality Test: Evaluates if an input is a 'typical' sample from the training distribution by analyzing its raw pixel statistics, such as compressibility or high-frequency artifacts, which often reveal adversarial or anomalous origins.

These methods quantify the model's epistemic uncertainty—the uncertainty in the model's parameters—rather than just the aleatoric noise in the data.

• Gradient Magnitude: Measures the norm of the gradient of the loss with respect to the input; OOD inputs typically exhibit larger gradients as they require significant model adaptation.
• Bayesian Neural Networks & MC Dropout: Use multiple stochastic forward passes to compute the variance of predictions. High variance across passes indicates the model has not converged on a stable representation, signaling an OOD input.

A proactive defense that fine-tunes the model on an auxiliary dataset of proxy OOD examples to explicitly teach the model a decision boundary between ID and OOD.

• Training Objective: The model is trained to predict a uniform distribution over known classes for OOD inputs while maintaining high accuracy on ID data.
• Effectiveness: This method dramatically improves detection performance by giving the model a concrete signal of what 'unknown' looks like, rather than relying on implicit heuristics.

A classification paradigm where the model must simultaneously identify known classes and reject unknown classes not seen during training. Unlike closed-set classification, open set recognition explicitly models the boundary between known and unknown data.

• Differs from OOD detection by operating at the semantic class level rather than the distribution level
• Uses OpenMax or extreme value theory to calibrate rejection thresholds
• Critical for autonomous systems encountering novel objects in the wild

A parametric OOD detection method that computes the distance between a test sample and the closest class-conditional Gaussian distribution in feature space. Samples far from all class centroids are flagged as OOD.

• Operates on penultimate layer features rather than logits
• Requires fitting a class-conditional mean and covariance on training data
• Provides a mathematically principled confidence score with known statistical properties

Uses the Helmholtz free energy formulation to derive a scalar score from logits that separates in-distribution from OOD samples. Lower energy indicates higher likelihood of being in-distribution.

• Computationally efficient: requires only a single forward pass
• Outperforms softmax confidence on standard benchmarks
• Can be combined with temperature scaling for fine-grained calibration

A preprocessing-based OOD detector that applies temperature scaling and input perturbation to amplify the separation between in-distribution and OOD softmax scores.

• Adds a small gradient-based perturbation to push inputs toward higher confidence
• Temperature parameter controls the sharpness of the softmax distribution
• Simple to implement as a wrapper around any pre-trained classifier

A Bayesian-inspired approach that trains multiple models with different random initializations and uses the variance or entropy of their predictions as an OOD signal. High disagreement among ensemble members indicates epistemic uncertainty.

• Captures model uncertainty distinct from data uncertainty
• Naturally provides both prediction and uncertainty estimates
• More robust than single-model methods but incurs N× computational cost

Addresses a fundamental flaw in using raw likelihood for OOD detection: generative models can assign high likelihood to OOD inputs. Likelihood ratio methods correct this by comparing a semantic model against a background model.

• Background model captures low-level statistics (e.g., pixel correlations)
• Semantic model captures high-level structure
• The ratio isolates semantic novelty from statistical typicality