Refine
Year of publication
- 2022 (10) (remove)
Document Type
- Conference Proceeding (10) (remove)
Conference Type
- Konferenzartikel (10)
Language
- English (10)
Has Fulltext
- no (10)
Is part of the Bibliography
- yes (10)
Keywords
- Machine Learning (4)
- Robustness (4)
- Adversarial Attacks (1)
- Adversarial Robustness (1)
- Aliasing (1)
- CNN (1)
- CNNs (1)
- Computer Vision (1)
- Model Calibration (1)
- Nyquist-Shannon (1)
- Sampling (1)
- adversarial (1)
- autoML (1)
- autoattack (1)
- deep learning (1)
- lid (1)
- mahalanobis (1)
- neural architecture search (1)
- pruning (1)
- spectraldefense (1)
Institute
- IMLA - Institute for Machine Learning and Analytics (10) (remove)
Open Access
- Bronze (8)
- Open Access (8)
- Closed (2)
Deep learning models are intrinsically sensitive to distribution shifts in the input data. In particular, small, barely perceivable perturbations to the input data can force models to make wrong predictions with high confidence. An common defense mechanism is regularization through adversarial training which injects worst-case perturbations back into training to strengthen the decision boundaries, and to reduce overfitting. In this context, we perform an investigation of 3 × 3 convolution filters that form in adversarially- trained models. Filters are extracted from 71 public models of the ℓ ∞ -RobustBench CIFAR-10/100 and ImageNet1k leaderboard and compared to filters extracted from models built on the same architectures but trained without robust regularization. We observe that adversarially-robust models appear to form more diverse, less sparse, and more orthogonal convolution filters than their normal counterparts. The largest differences between robust and normal models are found in the deepest layers, and the very first convolution layer, which consistently and predominantly forms filters that can partially eliminate perturbations, irrespective of the architecture.
Many commonly well-performing convolutional neural network models have shown to be susceptible to input data perturbations, indicating a low model robustness. Adversarial attacks are thereby specifically optimized to reveal model weaknesses, by generating small, barely perceivable image perturbations that flip the model prediction. Robustness against attacks can be gained for example by using adversarial examples during training, which effectively reduces the measurable model attackability. In contrast, research on analyzing the source of a model’s vulnerability is scarce. In this paper, we analyze adversarially trained, robust models in the context of a specifically suspicious network operation, the downsampling layer, and provide evidence that robust models have learned to downsample more accurately and suffer significantly less from aliasing than baseline models.
Currently, many theoretical as well as practically relevant questions towards the transferability and robustness of Convolutional Neural Networks (CNNs) remain unsolved. While ongoing research efforts are engaging these problems from various angles, in most computer vision related cases these approaches can be generalized to investigations of the effects of distribution shifts in image data. In this context, we propose to study the shifts in the learned weights of trained CNN models. Here we focus on the properties of the distributions of dominantly used 3×3 convolution filter kernels. We collected and publicly provide a dataset with over 1.4 billion filters from hundreds of trained CNNs, using a wide range of datasets, architectures, and vision tasks. In a first use case of the proposed dataset, we can show highly relevant properties of many publicly available pre-trained models for practical applications: I) We analyze distribution shifts (or the lack thereof) between trained filters along different axes of meta-parameters, like visual category of the dataset, task, architecture, or layer depth. Based on these results, we conclude that model pre-training can succeed on arbitrary datasets if they meet size and variance conditions. II) We show that many pre-trained models contain degenerated filters which make them less robust and less suitable for fine-tuning on target applications. Data & Project website: https://github.com/paulgavrikov/cnn-filter-db.
Recent work has investigated the distributions of learned convolution filters through a large-scale study containing hundreds of heterogeneous image models. Surprisingly, on average, the distributions only show minor drifts in comparisons of various studied dimensions including the learned task, image domain, or dataset. However, among the studied image domains, medical imaging models appeared to show significant outliers through "spikey" distributions, and, therefore, learn clusters of highly specific filters different from other domains. Following this observation, we study the collected medical imaging models in more detail. We show that instead of fundamental differences, the outliers are due to specific processing in some architectures. Quite the contrary, for standardized architectures, we find that models trained on medical data do not significantly differ in their filter distributions from similar architectures trained on data from other domains. Our conclusions reinforce previous hypotheses stating that pre-training of imaging models can be done with any kind of diverse image data.
Estimating the Robustness of Classification Models by the Structure of the Learned Feature-Space
(2022)
Over the last decade, the development of deep image classification networks has mostly been driven by the search for the best performance in terms of classification accuracy on standardized benchmarks like ImageNet. More recently, this focus has been expanded by the notion of model robustness, \ie the generalization abilities of models towards previously unseen changes in the data distribution. While new benchmarks, like ImageNet-C, have been introduced to measure robustness properties, we argue that fixed testsets are only able to capture a small portion of possible data variations and are thus limited and prone to generate new overfitted solutions. To overcome these drawbacks, we suggest to estimate the robustness of a model directly from the structure of its learned feature-space. We introduce robustness indicators which are obtained via unsupervised clustering of latent representations from a trained classifier and show very high correlations to the model performance on corrupted test data.
Over the last years, Convolutional Neural Networks (CNNs) have been the dominating neural architecture in a wide range of computer vision tasks. From an image and signal processing point of view, this success might be a bit surprising as the inherent spatial pyramid design of most CNNs is apparently violating basic signal processing laws, i.e. Sampling Theorem in their down-sampling operations. However, since poor sampling appeared not to affect model accuracy, this issue has been broadly neglected until model robustness started to receive more attention. Recent work in the context of adversarial attacks and distribution shifts, showed after all, that there is a strong correlation between the vulnerability of CNNs and aliasing artifacts induced by poor down-sampling operations. This paper builds on these findings and introduces an aliasing free down-sampling operation which can easily be plugged into any CNN architecture: FrequencyLowCut pooling. Our experiments show, that in combination with simple and Fast Gradient Sign Method (FGSM) adversarial training, our hyper-parameter free operator substantially improves model robustness and avoids catastrophic overfitting. Our code is available at https://github.com/GeJulia/flc_pooling
In this paper, we propose a unified approach for network pruning and one-shot neural architecture search (NAS) via group sparsity. We first show that group sparsity via the recent Proximal Stochastic Gradient Descent (ProxSGD) algorithm achieves new state-of-the-art results for filter pruning. Then, we extend this approach to operation pruning, directly yielding a gradient-based NAS method based on group sparsity. Compared to existing gradient-based algorithms such as DARTS, the advantages of this new group sparsity approach are threefold. Firstly, instead of a costly bilevel optimization problem, we formulate the NAS problem as a single-level optimization problem, which can be optimally and efficiently solved using ProxSGD with convergence guarantees. Secondly, due to the operation-level sparsity, discretizing the network architecture by pruning less important operations can be safely done without any performance degradation. Thirdly, the proposed approach finds architectures that are both stable and well-performing on a variety of search spaces and datasets.
Recently, RobustBench (Croce et al. 2020) has become a widely recognized benchmark for the adversarial robustness of image
classification networks. In it’s most commonly reported sub-task, RobustBench evaluates and ranks the adversarial robustness of trained neural networks on CIFAR10 under AutoAttack (Croce and Hein 2020b) with l∞ perturbations limited to ϵ = 8/255. With leading scores of the currently best performing models of around 60% of the baseline, it is fair to characterize this benchmark to be quite challenging. Despite it’s general acceptance in recent literature, we aim to foster discussion about the suitability of RobustBench as a key indicator for robustness which could be generalized to practical applications. Our line of argumentation against this is two-fold and supported by excessive experiments presented in this paper: We argue that I) the alternation of data by AutoAttack with l∞, ϵ = 8/255 is unrealistically strong, resulting in close to perfect detection rates of adversarial samples even by simple detection algorithms and human observers.
We also show that other attack methods are much harder to detect while achieving similar success rates. II) That results on low resolution data sets like CIFAR10 do not generalize well to higher resolution images as gradient based attacks appear to become even more detectable with increasing resolutions.
Despite the success of convolutional neural networks (CNNs) in many academic benchmarks for computer vision tasks, their application in the real-world is still facing fundamental challenges. One of these open problems is the inherent lack of robustness, unveiled by the striking effectiveness of adversarial attacks. Current attack methods are able to manipulate the network's prediction by adding specific but small amounts of noise to the input. In turn, adversarial training (AT) aims to achieve robustness against such attacks and ideally a better model generalization ability by including adversarial samples in the trainingset. However, an in-depth analysis of the resulting robust models beyond adversarial robustness is still pending. In this paper, we empirically analyze a variety of adversarially trained models that achieve high robust accuracies when facing state-of-the-art attacks and we show that AT has an interesting side-effect: it leads to models that are significantly less overconfident with their decisions, even on clean data than non-robust models. Further, our analysis of robust models shows that not only AT but also the model's building blocks (like activation functions and pooling) have a strong influence on the models' prediction confidences. Data & Project website: https://github.com/GeJulia/robustness_confidences_evaluation
Despite the success of convolutional neural networks (CNNs) in many academic benchmarks for computer vision tasks, their application in the real-world is still facing fundamental challenges. One of these open problems is the inherent lack of robustness, unveiled by the striking effectiveness of adversarial attacks. Adversarial training (AT) is often considered as a remedy to train more robust networks. In this paper, we empirically analyze a variety of adversarially trained models that achieve high robust accuracies when facing state-of-the-art attacks and we show that AT has an interesting side-effect: it leads to models that are significantly less overconfident with their decisions even on clean data than non-robust models. Further, our analysis of robust models shows that not only AT but also the model's building blocks (like activation functions and pooling) have a strong influence on the models' prediction confidences.