Déjà Vu: an empirical evaluation of the memorization properties of ConvNets

Abstract

Convolutional neural networks memorize part of their training data, which is why strategies such as data augmentation and drop-out are employed to mitigate overfitting. This paper considers the related question of "membership inference", where the goal is to determine if an image was used during training. We consider it under three complementary angles. We show how to detect which dataset was used to train a model, and in particular whether some validation images were used at train time. We then analyze explicit memorization and extend classical random label experiments to the problem of learning a model that predicts if an image belongs to an arbitrary set. Finally, we propose a new approach to infer membership when a few of the top layers are not available or have been fine-tuned, and show that lower layers still carry information about the training samples. To support our findings, we conduct large-scale experiments on Imagenet and subsets of YFCC-100M with modern architectures such as VGG and Resnet.

References in corpus (10)

• Very Deep Convolutional Networks for Large-Scale Image Recognition
• YFCC100M: The New Data in Multimedia Research
• Understanding deep learning requires rethinking generalization
• Deep Image Prior
• The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networks
• Unsupervised Learning by Predicting Noise
• Understanding Membership Inferences on Well-Generalized Learning Models
• On-Average KL-Privacy and its equivalence to Generalization for Max-Entropy Mechanisms
• LOGAN: Membership Inference Attacks Against Generative Models
• ADA: A Game-Theoretic Perspective on Data Augmentation for Object Detection

Cited by in corpus (4)

• P3SGD: Patient Privacy Preserving SGD for Regularizing Deep CNNs in Pathological Image Classification
• Shift-Table: A Low-latency Learned Index for Range Queries using Model Correction
• Have you forgotten? A method to assess if machine learning models have forgotten data
• How BPE Affects Memorization in Transformers