[论文解读] Generative Probabilistic Novelty Detection with Adversarial Autoencoders

Stanislav Pidhorskyi, Ranya Almohsen+2 more|arXiv (Cornell University)|Jul 6, 2018

Anomaly Detection Techniques and Applications参考文献 50被引用 140

一句话总结

GPND 使用对抗自编码器将内点分布学习为参数化流形，并计算生成内点概率以检测新颖性，在若干基准测试上达到最先进的结果。

ABSTRACT

Novelty detection is the problem of identifying whether a new data point is considered to be an inlier or an outlier. We assume that training data is available to describe only the inlier distribution. Recent approaches primarily leverage deep encoder-decoder network architectures to compute a reconstruction error that is used to either compute a novelty score or to train a one-class classifier. While we too leverage a novel network of that kind, we take a probabilistic approach and effectively compute how likely is that a sample was generated by the inlier distribution. We achieve this with two main contributions. First, we make the computation of the novelty probability feasible because we linearize the parameterized manifold capturing the underlying structure of the inlier distribution, and show how the probability factorizes and can be computed with respect to local coordinates of the manifold tangent space. Second, we improved the training of the autoencoder network. An extensive set of results show that the approach achieves state-of-the-art results on several benchmark datasets.

研究动机与目标

• Motivate novelty detection when only inlier data is available.
• Model the inlier distribution as a parameterized manifold learned by an encoder-decoder network.
• Compute the inlier probability by linearizing the manifold and factorizing the distribution along tangent and noise components.
• Enhance training with two discriminators to regularize latent space and decoder output for better generation and detection.
• Demonstrate state-of-the-art performance across multiple image datasets and evaluation metrics.

提出的方法

• Represent inlier data as x = f(z) + ξ, with z in a latent space and f modeling a manifold M.
• Linearize f at the projected point to obtain a tangent space T and perform coordinate rotation to W parallel and W perpendicular components.
• Assume independence pX(x) = pW∥(w∥) pW⊥(w⊥) to derive a computable novelty test pX(x̄) ≥ γ for inliers and < γ for outliers.
• Learn f and g (encoder/decoder) via an Adversarial Autoencoder with two discriminators: Dz for latent space prior and Dx for decoder output realism, plus an autoencoder reconstruction loss.
• Estimate pZ(z) offline after training and model p‖W⊥‖ via histogram of norms; combine components to obtain pX(x̄) as the novelty score.

实验结果

研究问题

• RQ1 How can novelty detection be cast as probability evaluation under a learned inlier distribution rather than a reconstruction error-based score?
• RQ2 Can a tangent-space (local linear) approximation of the inlier manifold enable feasible computation of the inlier probability for new samples?
• RQ3 Does incorporating adversarial regularization on both latent space and decoded outputs improve the quality of the learned inlier manifold and the accuracy of novelty detection?
• RQ4 How does GPND perform across standard image datasets compared to existing state-of-the-art methods under various inlier/outlier settings?

主要发现

• GPND achieves strong, often state-of-the-art, performance across MNIST, Coil-100, Fashion-MNIST, CIFAR-10/100, and other benchmarks.
• The approach consistently improves over reconstruction-error-based methods and one-class classifiers by modeling the full inlier probability distribution.
• The dual-adversarial training (latent space and output distribution) plus the autoencoder loss yields better manifold fidelity and sharper novelty discrimination.
• Afeasible computation of novelty scores is achieved through local linearization of the manifold and independence of tangent/noise coordinates, enabling practical testing times.
• Compared with ODIN, GPND is competitive and can surpass OdIN on high-class-count datasets without requiring label information.