[Paper Review] Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networks

Shiyu Liang, Yixuan Li+1 more|arXiv (Cornell University)|Jun 8, 2017

Adversarial Robustness in Machine Learning46 references648 citations

TL;DR

ODIN improves out-of-distribution detection for pre-trained networks using temperature scaling and input perturbations, without retraining, achieving state-of-the-art results across multiple architectures and datasets.

ABSTRACT

We consider the problem of detecting out-of-distribution images in neural networks. We propose ODIN, a simple and effective method that does not require any change to a pre-trained neural network. Our method is based on the observation that using temperature scaling and adding small perturbations to the input can separate the softmax score distributions between in- and out-of-distribution images, allowing for more effective detection. We show in a series of experiments that ODIN is compatible with diverse network architectures and datasets. It consistently outperforms the baseline approach by a large margin, establishing a new state-of-the-art performance on this task. For example, ODIN reduces the false positive rate from the baseline 34.7% to 4.3% on the DenseNet (applied to CIFAR-10) when the true positive rate is 95%.

Motivation & Objective

• Motivate the need for reliable OOD detection in deployment of neural networks.
• Propose a no-retraining-detector (ODIN) that separates in- and out-of-distribution softmax scores.
• Evaluate ODIN across diverse architectures and datasets to establish robustness and improvement over baselines.
• Analyze how temperature, perturbation, and dataset distance affect detection performance.

Proposed method

• Apply temperature scaling to softmax outputs with parameter T.
• Add small input perturbations in the direction of the input gradient to amplify separability.
• Compute a calibrated softmax score on perturbed inputs and threshold to detect OOD samples.
• Tune T, perturbation magnitude epsilon, and threshold delta using a separate OOD validation set.
• Evaluate performance with metrics including FPR at 95% TPR, Detection Error, AUROC, and AUPR.

Experimental results

Research questions

• RQ1 Can a post-hoc detector without retraining improve OOD detection across common architectures?
• RQ2 How do temperature scaling and input perturbations interact to enhance separation between in- and out-of-distribution samples?
• RQ3 What is the impact of hyperparameters (T, epsilon) on detection performance across datasets?
• RQ4 How transferable are tuned parameters across different OOD datasets?

Key findings

• ODIN substantially improves OOD detection over the baseline, e.g., reducing FPR from 34.7% to 4.3% at 95% TPR on DenseNet-CIFAR-10 vs TinyImageNet (crop).
• ODIN maintains strong performance across DenseNet-BC-100 and Wide ResNet architectures on multiple OOD datasets (TinyImageNet, LSUN, Gaussian/Uniform noise).
• Using a large temperature T (e.g., 1000) combined with carefully chosen perturbation epsilon significantly boosts separability between in- and out-of-distribution examples.
• Hyperparameters tuned on one validation OOD set generalize well to other OOD test sets with similar performance, indicating parameter transferability.
• As OOD dataset distance (MMD) increases, ODIN’s performance tends to improve, while closer distributions make detection harder.