ICLR 2023 Papers

Adversarial training suffers from the issue of robust overfitting, which seriously impairs its generalization performance. Data augmentation, which is effective at preventing overfitting in standard training, has been observed by many previous works to be ineffective in mitigating overfitting in adversarial training. This work proves that, contrary to previous findings, data augmentation alone can significantly boost accuracy and robustness in adversarial training. We find that the hardness and the diversity of data augmentation are important factors in combating robust overfitting. In general, diversity can improve both accuracy and robustness, while hardness can boost robustness at the cost of accuracy within a certain limit and degrade them both over that limit. To mitigate robust overfitting, we first propose a new crop transformation Cropshift with improved diversity compared to the conventional one (Padcrop). We then propose a new data augmentation scheme, based on Cropshift, with much improved diversity and well-balanced hardness. Empirically, our augmentation method achieves the state-of-the-art accuracy and robustness for data augmentations in adversarial training. Furthermore, it matches, or even exceeds when combined with weight averaging, the performance of the best contemporary regularization methods for alleviating robust overfitting.

Ensembling certifiably robust neural networks is a promising approach for improving the \emph{certified robust accuracy} of neural models. Black-box ensembles that assume only query-access to the constituent models (and their robustness certifiers) during prediction are particularly attractive due to their modular structure. Cascading ensembles are a popular instance of black-box ensembles that appear to improve certified robust accuracies in practice. However, we show that the robustness certifier used by a cascading ensemble is unsound. That is, when a cascading ensemble is certified as locally robust at an input $x$ (with respect to $\epsilon$), there can be inputs $x'$ in the $\epsilon$-ball centered at $x$, such that the cascade's prediction at $x'$ is different from $x$ and thus the ensemble is not locally robust. Our theoretical findings are accompanied by empirical results that further demonstrate this unsoundness. We present a new attack against cascading ensembles and show that: (1) there exists an adversarial input for up to 88\% of the samples where the ensemble claims to be certifiably robust and accurate; and (2) the accuracy of a cascading ensemble under our attack is as low as 11\% when it claims to be certifiably robust and accurate on 97\% of the test set. Our work reveals a critical pitfall of cascading certifiably robust models by showing that the seemingly beneficial strategy of cascading can actually hurt the robustness of the resulting ensemble. Our code is available at https://github.com/TristaChi/ensembleKW.

As machine learning models are increasingly being employed to make consequential decisions in real-world settings, it becomes critical to ensure that individuals who are adversely impacted (e.g., loan denied) by the predictions of these models are provided with a means for recourse. While several approaches have been proposed to construct recourses for affected individuals, the recourses output by these methods either achieve low costs (i.e., ease-of-implementation) or robustness to small perturbations (i.e., noisy implementations of recourses), but not both due to the inherent trade-offs between the recourse costs and robustness. Furthermore, prior approaches do not provide end users with any agency over navigating the aforementioned trade-offs. In this work, we address the above challenges by proposing the first algorithmic framework which enables users to effectively manage the recourse cost vs. robustness trade-offs. More specifically, our framework Probabilistically ROBust rEcourse (PROBE) lets users choose the probability with which a recourse could get invalidated (recourse invalidation rate) if small changes are made to the recourse i.e., the recourse is implemented somewhat noisily. To this end, we propose a novel objective function which simultaneously minimizes the gap between the achieved (resulting) and desired recourse invalidation rates, minimizes recourse costs, and also ensures that the resulting recourse achieves a positive model prediction. We develop novel theoretical results to characterize the recourse invalidation rates corresponding to any given instance w.r.t. different classes of underlying models (e.g., linear models, tree based models etc.), and leverage these results to efficiently optimize the proposed objective. Experimental evaluation with multiple real world datasets demonstrates the efficacy of the proposed framework.

Neural networks can be drastically shrunk in size by removing redundant parameters. While crucial for the deployment on resource-constraint hardware, oftentimes, compression comes with a severe drop in accuracy and lack of adversarial robustness. Despite recent advances, counteracting both aspects has only succeeded for moderate compression rates so far. We propose a novel method, HARP, that copes with aggressive pruning significantly better than prior work. For this, we consider the network holistically. We learn a global compression strategy that optimizes how many parameters (compression rate) and which parameters (scoring connections) to prune specific to each layer individually. Our method fine-tunes an existing model with dynamic regularization, that follows a step-wise incremental function balancing the different objectives. It starts by favoring robustness before shifting focus on reaching the target compression rate and only then handles the objectives equally. The learned compression strategies allow us to maintain the pre-trained model’s natural accuracy and its adversarial robustness for a reduction by 99% of the network’s original size. Moreover, we observe a crucial influence of non-uniform compression across layers. The implementation of HARP is publicly available at https://intellisec.de/research/harp.

Despite their verified performance, neural networks are prone to be misled by maliciously designed adversarial examples. This work investigates the robustness of neural networks from the activation pattern perspective. We find that despite the complex structure of the deep neural network, most of the neurons provide locally stable contributions to the output, while the minority, which we refer to as float neurons, can greatly affect the prediction. We decompose the computational graph of the neural network into the fixed path and float path and investigate their role in generating adversarial examples. Based on our analysis, we categorize the vulnerable examples into Lipschitz vulnerability and float neuron vulnerability. We show that the boost of robust accuracy from randomized smoothing is the result of correcting the latter. We then propose an SC-RFP (smoothed classifier with repressed float path) to further reduce the instability of the float neurons and show that our result can provide a higher certified radius as well as accuracy.

We develop ShiftMatch, a new training-data-dependent likelihood for robustness to corruption in Bayesian neural networks (BNNs). ShiftMatch is inspired by the training-data-dependent “EmpCov” priors from Izmailov et al. (2021a), and efficiently matches test-time spatial correlations to those at training time. Critically, ShiftMatch is designed to leave the neural network’s training time likelihood unchanged, allowing it to use publicly available samples from pre-trained BNNs. Using pre-trained HMC samples, ShiftMatch gives strong performance improvements on CIFAR-10-C, outperforms EmpCov priors (though ShiftMatch uses extra information from a minibatch of corrupted test points), and is perhaps the first Bayesian method capable of convincingly outperforming plain deep ensembles.

Important research efforts have focused on the design and training of neural networks with a controlled Lipschitz constant. The goal is to increase and sometimes guarantee the robustness against adversarial attacks. Recent promising techniques draw inspirations from different backgrounds to design 1-Lipschitz neural networks, just to name a few: convex potential layers derive from the discretization of continuous dynamical systems, Almost-Orthogonal-Layer proposes a tailored method for matrix rescaling. However, it is today important to consider the recent and promising contributions in the field under a common theoretical lens to better design new and improved layers. This paper introduces a novel algebraic perspective unifying various types of 1-Lipschitz neural networks, including the ones previously mentioned, along with methods based on orthogonality and spectral methods. Interestingly, we show that many existing techniques can be derived and generalized via finding analytical solutions of a common semidefinite programming (SDP) condition. We also prove that AOL biases the scaled weight to the ones which are close to the set of orthogonal matrices in a certain mathematical manner. Moreover, our algebraic condition, combined with the Gershgorin circle theorem, readily leads to new and diverse parameterizations for 1-Lipschitz network layers. Our approach, called SDP-based Lipschitz Layers (SLL), allows us to design non-trivial yet efficient generalization of convex potential layers. Finally, the comprehensive set of experiments on image classification shows that SLLs outperform previous approaches on certified robust accuracy. Code is available at https://github.com/araujoalexandre/Lipschitz-SLL-Networks.

Neural network classifiers can largely rely on simple spurious features, such as image backgrounds, to make predictions. However, even in these cases, we show that they still often learn core features associated with the desired attributes of the data, contrary to recent findings. Inspired by this insight, we demonstrate that simple last layer retraining can match or outperform state-of-the-art approaches on spurious correlation benchmarks, but with profoundly lower complexity and computational expenses. Moreover, we show that last layer retraining on large ImageNet-trained models can also significantly reduce reliance on background and texture information, improving robustness to covariate shift, after only minutes of training on a single GPU.

The vulnerability of deep neural networks (DNNs) to adversarial examples has attracted great attention in the machine learning community. The problem is related to non-flatness and non-smoothness of normally obtained loss landscapes. Training augmented with adversarial examples (a.k.a., adversarial training) is considered as an effective remedy. In this paper, we highlight that some collaborative examples, nearly perceptually indistinguishable from both adversarial and benign examples yet show extremely lower prediction loss, can be utilized to enhance adversarial training. A novel method is therefore proposed to achieve new state-of-the-arts in adversarial robustness. Code: https://github.com/qizhangli/ST-AT.

Reliable application of machine learning-based decision systems in the wild is one of the major challenges currently investigated by the field. A large portion of established approaches aims to detect erroneous predictions by means of assigning confidence scores. This confidence may be obtained by either quantifying the model's predictive uncertainty, learning explicit scoring functions, or assessing whether the input is in line with the training distribution. Curiously, while these approaches all state to address the same eventual goal of detecting failures of a classifier upon real-world application, they currently constitute largely separated research fields with individual evaluation protocols, which either exclude a substantial part of relevant methods or ignore large parts of relevant failure sources. In this work, we systematically reveal current pitfalls caused by these inconsistencies and derive requirements for a holistic and realistic evaluation of failure detection. To demonstrate the relevance of this unified perspective, we present a large-scale empirical study for the first time enabling benchmarking confidence scoring functions w.r.t all relevant methods and failure sources. The revelation of a simple softmax response baseline as the overall best performing method underlines the drastic shortcomings of current evaluation in the plethora of publicized research on confidence scoring. Code and trained models are at https://github.com/https://github.com/IML-DKFZ/fd-shifts

Neural text-to-SQL models have achieved remarkable performance in translating natural language questions into SQL queries. However, recent studies reveal that text-to-SQL models are vulnerable to task-specific perturbations. Previous curated robustness test sets usually focus on individual phenomena. In this paper, we propose a comprehensive robustness benchmark based on Spider, a cross-domain text-to-SQL benchmark, to diagnose the model robustness. We design 17 perturbations on databases, natural language questions, and SQL queries to measure the robustness from different angles. In order to collect more diversified natural question perturbations, we utilize large pretrained language models (PLMs) to simulate human behaviors in creating natural questions. We conduct a diagnostic study of the state-of-the-art models on the robustness set. Experimental results reveal that even the most robust model suffers from a 14.0% performance drop overall and a 50.7% performance drop on the most challenging perturbation. We also present a breakdown analysis regarding text-to-SQL model designs and provide insights for improving model robustness.

Geometric image transformations that arise in the real world, such as scaling and rotation, have been shown to easily deceive deep neural networks (DNNs). Hence, training DNNs to be certifiably robust to these perturbations is critical. However, no prior work has been able to incorporate the objective of deterministic certified robustness against geometric transformations into the training procedure, as existing verifiers are exceedingly slow. To address these challenges, we propose the first provable defense for deterministic certified geometric robustness. Our framework leverages a novel GPU-optimized verifier that can certify images between 60$\times$ to 42,600$\times$ faster than existing geometric robustness verifiers, and thus unlike existing works, is fast enough for use in training. Across multiple datasets, our results show that networks trained via our framework consistently achieve state-of-the-art deterministic certified geometric robustness and clean accuracy. Furthermore, for the first time, we verify the geometric robustness of a neural network for the challenging, real-world setting of autonomous driving.

The recent success of Vision Transformers is shaking the long dominance of Convolutional Neural Networks (CNNs) in image recognition for a decade. Specifically, in terms of robustness on out-of-distribution samples, recent research finds that Transformers are inherently more robust than CNNs, regardless of different training setups. Moreover, it is believed that such superiority of Transformers should largely be credited to their \emph{self-attention-like architectures per se}. In this paper, we question that belief by closely examining the design of Transformers. Our findings lead to three highly effective architecture designs for boosting robustness, yet simple enough to be implemented in several lines of code, namely a) patchifying input images, b) enlarging kernel size, and c) reducing activation layers and normalization layers. Bringing these components together, we are able to build pure CNN architectures without any attention-like operations that are as robust as, or even more robust than, Transformers. We hope this work can help the community better understand the design of robust neural architectures. The code is publicly available at https://github.com/UCSC-VLAA/RobustCNN.

This paper studies a curious phenomenon that machine learning model with Transformer architectures have sparse activation maps. By activation map we refer to the intermediate output of the multi-layer perceptrons (MLPs) after a ReLU activation function, and by "sparse" we mean that on average very few entries (e.g., 3.0% for T5-Base and 6.3% for ViT-B16) are nonzero for each input to MLP. Moreover, larger Transformers with more layers and wider MLP hidden dimensions are sparser as measured by the percentage of nonzero entries. Through extensive experiments we demonstrate that the emergence of sparsity is a prevalent phenomenon that occurs for both natural language processing and vision tasks, on both training and evaluation data, for Transformers of various configurations, at layers of all depth levels. We discuss how sparsity immediately implies a way to significantly reduce the FLOP count and improve efficiency for Transformers. Moreover, we demonstrate perhaps surprisingly that enforcing an even sparser activation via Top-k thresholding with a small k brings a collection of desired properties, namely less sensitivity to noisy training data, more robustness to input corruptions, and better calibration for their prediction confidence.

Real-world machine learning problems often exhibit shifts between the source and target distributions, in which source data does not fully convey the desired behavior on target inputs. Different functions that achieve near-perfect source accuracy can make differing predictions on test inputs, and such ambiguity makes robustness to distribution shifts challenging. We propose DivDis, a simple two-stage framework for identifying and resolving ambiguity in data. DivDis first learns a diverse set of hypotheses that achieve low source loss but make differing predictions on target inputs. We then disambiguate by selecting one of the discovered functions using additional information, for example, a small number of target labels. Our experimental evaluation shows improved performance in subpopulation shift and domain generalization settings, demonstrating that DivDis can scalably adapt to distribution shifts in image and text classification benchmarks.

Standard empirical risk minimization (ERM) training can produce deep neural network (DNN) models that are accurate on average but under-perform in under-represented population subgroups, especially when there are imbalanced group distributions in the long-tailed training data. Therefore, approaches that improve the accuracy - group robustness trade-off frontier of a DNN model (i.e. improving worst-group accuracy without sacrificing average accuracy, or vice versa) is of crucial importance. Uncertainty-based active learning (AL) can potentially improve the frontier by preferentially sampling underrepresented subgroups to create a more balanced training dataset. However, the quality of uncertainty estimates from modern DNNs tend to degrade in the presence of spurious correlations and dataset bias, compromising the effectiveness of AL for sampling tail groups. In this work, we propose Introspective Self-play (ISP), a simple approach to improve the uncertainty estimation of a deep neural network under dataset bias, by adding an auxiliary introspection task requiring a model to predict the bias for each data point in addition to the label. We show that ISP provably improves the bias-awareness of the model representation and the resulting uncertainty estimates. On two real-world tabular and language tasks,ISP serves as a simple “plug-in” for AL model training, consistently improving both the tail-group sampling rate and the final accuracy-fairness trade-off frontier of popular AL methods.

In this work, we explore the maximum-margin bias of quasi-homogeneous neural networks trained with gradient flow on an exponential loss and past a point of separability. We introduce the class of quasi-homogeneous models, which is expressive enough to describe nearly all neural networks with homogeneous activations, even those with biases, residual connections, and normalization layers, while structured enough to enable geometric analysis of its gradient dynamics. Using this analysis, we generalize the existing results of maximum-margin bias for homogeneous networks to this richer class of models. We find that gradient flow implicitly favors a subset of the parameters, unlike in the case of a homogeneous model where all parameters are treated equally. We demonstrate through simple examples how this strong favoritism toward minimizing an asymmetric norm can degrade the robustness of quasi-homogeneous models. On the other hand, we conjecture that this norm-minimization discards, when possible, unnecessary higher-order parameters, reducing the model to a sparser parameterization. Lastly, by applying our theorem to sufficiently expressive neural networks with normalization layers, we reveal a universal mechanism behind the empirical phenomenon of Neural Collapse.

Randomized smoothing is sound when using infinite precision. However, we show that randomized smoothing is no longer sound for limited floating-point precision. We present a simple example where randomized smoothing certifies a radius of $1.26$ around a point, even though there is an adversarial example in the distance $0.8$ and show how this can be abused to give false certificates for CIFAR10. We discuss the implicit assumptions of randomized smoothing and show that they do not apply to generic image classification models whose smoothed versions are commonly certified. In order to overcome this problem, we propose a sound approach to randomized smoothing when using floating-point precision with essentially equal speed for quantized input. It yields sound certificates or image classifiers which for the ones tested so far are very similar to the unsound practice of randomized smoothing. Our only assumption is that we have access to a fair coin.

While adversarial training is generally used as a defense mechanism, recent works show that it can also act as a regularizer. By co-training a neural network on clean and adversarial inputs, it is possible to improve classification accuracy on the clean, non-adversarial inputs. We demonstrate that, contrary to previous findings, it is not necessary to separate batch statistics when co-training on clean and adversarial inputs, and that it is sufficient to use adapters with few domain-specific parameters for each type of input. We establish that using the classification token of a Vision Transformer (ViT) as an adapter is enough to match the classification performance of dual normalization layers, while using significantly less additional parameters. First, we improve upon the top-1 accuracy of a non-adversarially trained ViT-B16 model by +1.12% on ImageNet (reaching 83.76% top-1 accuracy). Second, and more importantly, we show that training with adapters enables model soups through linear combinations of the clean and adversarial tokens. These model soups, which we call adversarial model soups, allow us to trade-off between clean and robust accuracy without sacrificing efficiency. Finally, we show that we can easily adapt the resulting models in the face of distribution shifts. Our ViT-B16 obtains top-1 accuracies on ImageNet variants that are on average +4.00% better than those obtained with Masked Autoencoders.