ICLR 2023 Papers

The use of well-disentangled representations offers many advantages for downstream tasks, e.g. an increased sample efficiency, or better interpretability. However, the quality of disentangled interpretations is often highly dependent on the choice of dataset-specific hyperparameters, in particular the regularization strength. To address this issue, we introduce DAVA, a novel training procedure for variational auto-encoders. DAVA completely alleviates the problem of hyperparameter selection. We compare DAVA to models with optimal hyperparameters. Without any hyperparameter tuning, DAVA is competitive on a diverse range of commonly used datasets. Underlying DAVA, we discover a necessary condition for unsupervised disentanglement, which we call PIPE. We demonstrate the ability of PIPE to positively predict the performance of downstream models in abstract reasoning. We also thoroughly investigate correlations with existing supervised and unsupervised metrics. The code is available at https://github.com/besterma/dava.

The ideally disentangled latent space in GAN involves the global representation of latent space using semantic attribute coordinates. In other words, in this disentangled space, there exists the global semantic basis as a vector space where each basis component describes one attribute of generated images. In this paper, we propose an unsupervised method for finding this global semantic basis in the intermediate latent space in GANs. This semantic basis represents sample-independent meaningful perturbations that change the same semantic attribute of an image on the entire latent space. The proposed global basis, called Fréchet basis, is derived by introducing Fréchet mean to the local semantic perturbations in a latent space. Fréchet basis is discovered in two stages. First, the global semantic subspace is discovered by the Fréchet mean in the Grassmannian manifold of the local semantic subspaces. Second, Fréchet basis is found by optimizing a basis of the semantic subspace via the Fréchet mean in the Special Orthogonal Group. Experimental results demonstrate that Fréchet basis provides better semantic factorization and robustness compared to the previous methods. Moreover, we suggest the basis refinement scheme for the previous methods. The quantitative experiments show that the refined basis achieves better semantic factorization while constrained on the same semantic subspace given by the previous method.

Large language models (LLMs) show impressive abilities via few-shot prompting. Commercialized APIs such as OpenAI GPT-3 further increase their use in real-world language applications. However, the crucial problem of how to improve the reliability of GPT-3 is still under-explored. While reliability is a broad and vaguely defined term, we decompose reliability into four main facets that correspond to the existing framework of ML safety and are well-recognized to be important: generalizability, social biases, calibration, and factuality. Our core contribution is to establish simple and effective prompts that improve GPT-3’s reliability as it: 1) generalizes out-of-distribution, 2) balances demographic distribution and uses natural language instructions to reduce social biases, 3) calibrates output probabilities, and 4) updates the LLM’s factual knowledge and reasoning chains. With appropriate prompts, GPT-3 is more reliable than smaller-scale supervised models on all these facets. We release all processed datasets, evaluation scripts, and model predictions. Our systematic empirical study not only sheds new insights on the reliability of prompting LLMs, but more importantly, our prompting strategies can help practitioners more reliably use LLMs like GPT-3.

To obtain, deterministic guarantees of adversarial robustness, specialized training methods are used. We propose, SABR, a novel such certified training method, based on the key insight that propagating interval bounds for a small but carefully selected subset of the adversarial input region is sufficient to approximate the worst-case loss over the whole region while significantly reducing approximation errors. We show in an extensive empirical evaluation that SABR outperforms existing certified defenses in terms of both standard and certifiable accuracies across perturbation magnitudes and datasets, pointing to a new class of certified training methods promising to alleviate the robustness-accuracy trade-off.

Graph Neural Networks (GNNs) have achieved state-of-the-art results on a variety of graph learning tasks, however, it has been demonstrated that they are vulnerable to adversarial attacks, raising serious security concerns. A lot of studies have been developed to train GNNs in a noisy environment and increase their robustness against adversarial attacks. However, existing methods have not uncovered a principled difficulty: the convoluted mixture distribution between clean and attacked data samples, which leads to sub-optimal model design and limits their frameworks’ robustness. In this work, we first begin by identifying the root cause of mixture distribution, then, for tackling it, we propose a novel method GAME - Graph Adversarial Mixture of Experts to enlarge the model capacity and enrich the representation diversity of adversarial samples, from three perspectives of model, training, and optimization. Specifically, we first propose a plug-and- play GAME layer that can be easily incorporated into any GNNs and enhance their adversarial learning capabilities. Second, we design a decoupling-based graph adversarial training in which the component of the model used to generate adversarial graphs is separated from the component used to update weights. Third, we introduce a graph diversity regularization that enables the model to learn diverse representation and further improves model performance. Extensive experiments demonstrate the effectiveness and advantages of GAME over the state-of-the-art adversarial training methods across various datasets given different attacks.

We tackle the domain generalisation (DG) problem by posing it as a domain adaptation (DA) task where we adversarially synthesise the worst-case `target' domain and adapt a model to that worst-case domain, thereby improving the model’s robustness. To synthesise data that is challenging yet semantics-preserving, we generate Fourier amplitude images and combine them with source domain phase images, exploiting the widely believed conjecture from signal processing that amplitude spectra mainly determines image style, while phase data mainly captures image semantics. To synthesise a worst-case domain for adaptation, we train the classifier and the amplitude generator adversarially. Specifically, we exploit the maximum classifier discrepancy (MCD) principle from DA that relates the target domain performance to the discrepancy of classifiers in the model hypothesis space. By Bayesian hypothesis modeling, we express the model hypothesis space effectively as a posterior distribution over classifiers given the source domains, making adversarial MCD minimisation feasible. On the DomainBed benchmark including the large-scale DomainNet dataset, the proposed approach yields significantly improved domain generalisation performance over the state-of-the-art.

In this paper, we investigate the problem of certifying neural network robustness against universal perturbations (UPs), which have been widely used in universal adversarial attacks and backdoor attacks. Existing robustness certification methods aim to provide robustness guarantees for each sample with respect to the worst-case perturbations given a neural network. However, those sample-wise bounds will be loose when considering the UP threat model as they overlook the important constraint that the perturbation should be shared across all samples. We propose a method based on a combination of linear relaxation-based perturbation analysis and Mixed Integer Linear Programming to establish the first robust certification method for UP. In addition, we develop a theoretical framework for computing error bounds on the entire population using the certification results from a randomly sampled batch. Aside from an extensive evaluation of the proposed certification, we further show how the certification facilitates efficient comparison of robustness among different models or efficacy among different universal adversarial attack defenses and enables accurate detection of backdoor target classes.

Robustness evaluation against adversarial examples has become increasingly important to unveil the trustworthiness of the prevailing deep models in natural language processing (NLP). However, in contrast to the computer vision domain where the first-order projected gradient descent (PGD) is used as the benchmark approach to generate adversarial examples for robustness evaluation, there lacks a principled first-order gradient-based robustness evaluation framework in NLP. The emerging optimization challenges lie in 1) the discrete nature of textual inputs together with the strong coupling between the perturbation location and the actual content, and 2) the additional constraint that the perturbed text should be fluent and achieve a low perplexity under a language model. These challenges make the development of PGD-like NLP attacks difficult. To bridge the gap, we propose TextGrad, a new attack generator using gradient-driven optimization, supporting high-accuracy and high-quality assessment of adversarial robustness in NLP. Specifically, we address the aforementioned challenges in a unified optimization framework. And we develop an effective convex relaxation method to co-optimize the continuously-relaxed site selection and perturbation variables and leverage an effective sampling method to establish an accurate mapping from the continuous optimization variables to the discrete textual perturbations. Moreover, as a first-order attack generation method, TextGrad can be baked into adversarial training to further improve the robustness of NLP models. Extensive experiments are provided to demonstrate the effectiveness of TextGrad not only in attack generation for robustness evaluation but also in adversarial defense. From the attack perspective, we show that TextGrad achieves remarkable improvements in both the attack success rate and the perplexity score over five state-of-the-art baselines. From the defense perspective, TextGrad-enabled adversarial training yields the most robust NLP model against a wide spectrum of NLP attacks.

Deep learning models have been widely used in commercial acoustic systems in recent years. However, adversarial audio examples can cause abnormal behaviors for those acoustic systems, while being hard for humans to perceive. Various methods, such as transformation-based defenses and adversarial training, have been proposed to protect acoustic systems from adversarial attacks, but they are less effective against adaptive attacks. Furthermore, directly applying the methods from the image domain can lead to suboptimal results because of the unique properties of audio data. In this paper, we propose an adversarial purification-based defense pipeline, AudioPure, for acoustic systems via off-the-shelf diffusion models. Taking advantage of the strong generation ability of diffusion models, AudioPure first adds a small amount of noise to the adversarial audio and then runs the reverse sampling step to purify the noisy audio and recover clean audio. AudioPure is a plug-and-play method that can be directly applied to any pretrained classifier without any fine-tuning or re-training. We conduct extensive experiments on the speech command recognition task to evaluate the robustness of AudioPure. Our method is effective against diverse adversarial attacks (e.g. L2 or L∞-norm). It outperforms the existing methods under both strong adaptive white-box and black-box attacks bounded by L2 or L∞-norm (up to +20% in robust accuracy). Besides, we also evaluate the certified robustness for perturbations bounded by L2-norm via randomized smoothing. Our pipeline achieves a higher certified accuracy than baselines.

The appearance of adversarial examples raises attention from both academia and industry. Along with the attack-defense arms race, adversarial training is the most effective against adversarial examples. However, we find inequality phenomena occur during the $l_{\infty}$-adversarial training, that few features dominate the prediction made by the adversarially trained model. We systematically evaluate such inequality phenomena by extensive experiments and find such phenomena become more obvious when performing adversarial training with increasing adversarial strength (evaluated by $\epsilon$). We hypothesize such inequality phenomena make $l_{\infty}$-adversarially trained model less reliable than the standard trained model when few ``important features" are influenced. To validate our hypothesis, we proposed two simple attacks that either perturb or replace important features with noise or occlusion. Experiments show that $l_{\infty}$-adversarially trained model can be easily attacked when the few important features are influenced. Our work shed light on the limitation of the practicality of $l_{\infty}$-adversarial training.

Humans have been shown to use a ''straightened'' encoding to represent the natural visual world as it evolves in time (Henaff et al. 2019). In the context of discrete video sequences, ''straightened'' means that changes between frames follow a more linear path in representation space at progressively deeper levels of processing. While deep convolutional networks are often proposed as models of human visual processing, many do not straighten natural videos. In this paper, we explore the relationship between network architecture, differing types of robustness, biologically-inspired filtering mechanisms, and representational straightness in response to time-varying input; we identify strengths and limitations of straightness as a useful way of evaluating neural network representations. We find that (1) adversarial training leads to straighter representations in both CNN and transformer-based architectures but (2) this effect is task-dependent, not generalizing to tasks such as segmentation and frame-prediction, where straight representations are not favorable for predictions; and nor to other types of robustness. In addition, (3) straighter representations impart temporal stability to class predictions, even for out-of-distribution data. Finally, (4) biologically-inspired elements increase straightness in the early stages of a network, but do not guarantee increased straightness in downstream layers of CNNs. We show that straightness is an easily computed measure of representational robustness and stability, as well as a hallmark of human representations with benefits for computer vision models.

The generalization ability of machine learning models degrades significantly when the test distribution shifts away from the training distribution. We investigate the problem of training models that are robust to shifts caused by changes in the distribution of class-priors or group-priors. The presence of skewed training priors can often lead to the models overfitting to spurious features. Unlike existing methods, which optimize for either the worst or the average performance over classes or groups, our work is motivated by the need for finer control over the robustness properties of the model. We present an extremely lightweight post-hoc approach that performs scaling adjustments to predictions from a pre-trained model, with the goal of minimizing a distributionally robust loss around a chosen target distribution. These adjustments are computed by solving a constrained optimization problem on a validation set and applied to the model during test time. Our constrained optimization objective is inspired from a natural notion of robustness to controlled distribution shifts. Our method comes with provable guarantees and empirically makes a strong case for distributional robust post-hoc classifiers. An empirical implementation is available at https://github.com/weijiaheng/Drops.

Training machine learning models robust to distribution shifts is critical for real-world applications. Some robust training algorithms (e.g., Group DRO) specialize to group shifts and require group information on all training points. Other methods (e.g., CVaR DRO) that do not need group annotations can be overly conservative, since they naively upweight high loss points which may form a contrived set that does not correspond to any meaningful group in the real world (e.g., when the high loss points are randomly mislabeled training points). In this work, we address limitations in prior approaches by assuming a more nuanced form of group shift: conditioned on the label, we assume that the true group function (indicator over group) is simple. For example, we may expect that group shifts occur along low bitrate features (e.g., image background, lighting). Thus, we aim to learn a model that maintains high accuracy on simple group functions realized by these low bitrate features, that need not spend valuable model capacity achieving high accuracy on contrived groups of examples. Based on this, we consider the two-player game formulation of DRO where the adversary's capacity is bitrate-constrained. Our resulting practical algorithm, Bitrate-Constrained DRO (\bdro), does not require group information on training samples yet matches the performance of Group DRO on datasets that have training group annotations and that of CVaR DRO on long-tailed distributions. Our theoretical analysis reveals that in some settings \bdro objective can provably yield statistically efficient and less conservative solutions than unconstrained CVaR DRO.

Recent works have shown that self-supervised learning can achieve remarkable robustness when integrated with adversarial training (AT). However, the robustness gap between supervised AT (sup-AT) and self-supervised AT (self-AT) remains significant. Motivated by this observation, we revisit existing self-AT methods and discover an inherent dilemma that affects self-AT robustness: either strong or weak data augmentations are harmful to self-AT, and a medium strength is insufficient to bridge the gap. To resolve this dilemma, we propose a simple remedy named DYNACL (Dynamic Adversarial Contrastive Learning). In particular, we propose an augmentation schedule that gradually anneals from a strong augmentation to a weak one to benefit from both extreme cases. Besides, we adopt a fast post-processing stage for adapting it to downstream tasks. Through extensive experiments, we show that DYNACL can improve state-of-the-art self-AT robustness by 8.84% under Auto-Attack on the CIFAR-10 dataset, and can even outperform vanilla supervised adversarial training for the first time. Our code is available at \url{https://github.com/PKU-ML/DYNACL}.

In supervised learning, it has been shown that label noise in the data can be interpolated without penalties on test accuracy. We show that interpolating label noise induces adversarial vulnerability, and prove the first theorem showing the relationship between label noise and adversarial risk for any data distribution. Our results are almost tight if we do not make any assumptions on the inductive bias of the learning algorithm. We then investigate how different components of this problem affect this result including properties of the distribution. We also discuss non-uniform label noise distributions; and prove a new theorem showing uniform label noise induces nearly as large an adversarial risk as the worst poisoning with the same noise rate. Then, we provide theoretical and empirical evidence that uniform label noise is more harmful than typical real-world label noise. Finally, we show how inductive biases amplify the effect of label noise and argue the need for future work in this direction.

Models for image segmentation, node classification and many other tasks map a single input to multiple labels. By perturbing this single shared input (e.g. the image) an adversary can manipulate several predictions (e.g. misclassify several pixels). Collective robustness certification is the task of provably bounding the number of robust predictions under this threat model. The only dedicated method that goes beyond certifying each output independently is limited to strictly local models, where each prediction is associated with a small receptive field. We propose a more general collective robustness certificate for all types of models. We further show that this approach is beneficial for the larger class of softly local models, where each output is dependent on the entire input but assigns different levels of importance to different input regions (e.g. based on their proximity in the image). The certificate is based on our novel localized randomized smoothing approach, where the random perturbation strength for different input regions is proportional to their importance for the outputs. Localized smoothing Pareto-dominates existing certificates on both image segmentation and node classification tasks, simultaneously offering higher accuracy and stronger certificates.

With more people publishing their personal data online, unauthorized data usage has become a serious concern. The unlearnable examples strategies have been introduced to prevent third parties from training on the data without permission. They add perturbations to the users’ data before publishing, so as to make the models trained on the perturbed published dataset invalidated. These perturbations have been generated for a specific training setting and a target dataset. However, their unlearnable effects significantly decrease when used in other training settings or datasets. To tackle this issue, we propose a novel unlearnable strategy based on Class-wise Separability Discriminant (CSD), which boosts the transferability of the unlearnable perturbations by enhancing the linear separability. Extensive experiments demonstrate the transferability of the unlearnable examples crafted by our proposed method across training settings and datasets.

Machine learning classifiers with high test accuracy often perform poorly under adversarial attacks. It is commonly believed that adversarial training alleviates this issue. In this paper, we demonstrate that, surprisingly, the opposite can be true for a natural class of perceptible perturbations --- even though adversarial training helps when enough data is available, it may in fact hurt robust generalization in the small sample size regime. We first prove this phenomenon for a high-dimensional linear classification setting with noiseless observations. Using intuitive insights from the proof, we could surprisingly find perturbations on standard image datasets for which this behavior persists. Specifically, it occurs for perceptible attacks that effectively reduce class information such as object occlusions or corruptions.