Identifying Interpretable Subspaces in Image Representations

We propose Automatic Feature Explanation using Contrasting Concepts (FALCON), an interpretability framework to explain features of image representations. For a target feature, FALCON captions its highly activating cropped images using a large captioning dataset (like LAION-400m) and a pre-trained vision-language model like CLIP. Each word among the captions is scored and ranked leading to a small number of shared, human-understandable concepts that closely describe the target feature. FALCON also applies contrastive interpretation using lowly activating (counterfactual) images, to eliminate spurious concepts. Although many existing approaches interpret features independently, we observe in state-of-the-art self-supervised and supervised models, that less than 20% of the representation space can be explained by individual features. We show that features in larger spaces become more interpretable when studied in groups and can be explained with high-order scoring concepts through FALCON. We discuss how extracted concepts can be used to explain and debug failures in downstream tasks. Finally, we present a technique to transfer concepts from one (explainable) representation space to another unseen representation space by learning a simple linear transformation.

On Practical Aspects of Aggregation Defenses against Data Poisoning Attacks

The increasing access to data poses both opportunities and risks in deep learning, as one can manipulate the behaviors of deep learning models with malicious training samples. Such attacks are known as data poisoning. Recent advances in defense strategies against data poisoning have highlighted the effectiveness of aggregation schemes in achieving state-of-the-art results in certified poisoning robustness. However, the practical implications of these approaches remain unclear. Here we focus on Deep Partition Aggregation, a representative aggregation defense, and assess its practical aspects, including efficiency, performance, and robustness. For evaluations, we use ImageNet resized to a resolution of 64 by 64 to enable evaluations at a larger scale than previous ones. Firstly, we demonstrate a simple yet practical approach to scaling base models, which improves the efficiency of training and inference for aggregation defenses. Secondly, we provide empirical evidence supporting the data-to-complexity ratio, i.e. the ratio between the data set size and sample complexity, as a practical estimation of the maximum number of base models that can be deployed while preserving accuracy. Last but not least, we point out how aggregation defenses boost poisoning robustness empirically through the poisoning overfitting phenomenon, which is the key underlying mechanism for the empirical poisoning robustness of aggregations. Overall, our findings provide valuable insights for practical implementations of aggregation defenses to mitigate the threat of data poisoning.

Diffused Redundancy in Pre-trained Representations

Representations learned by pre-training a neural network on a large dataset are increasingly used successfully to perform a variety of downstream tasks. In this work, we take a closer look at how features are encoded in such pre-trained representations. We find that learned representations in a given layer exhibit a degree of diffuse redundancy, i.e., any randomly chosen subset of neurons in the layer that is larger than a threshold size shares a large degree of similarity with the full layer and is able to perform similarly as the whole layer on a variety of downstream tasks. For example, a linear probe trained on $20\%$ of randomly picked neurons from a ResNet50 pre-trained on ImageNet1k achieves an accuracy within $5\%$ of a linear probe trained on the full layer of neurons for downstream CIFAR10 classification. We conduct experiments on different neural architectures (including CNNs and Transformers) pre-trained on both ImageNet1k and ImageNet21k and evaluate a variety of downstream tasks taken from the VTAB benchmark. We find that the loss & dataset used during pre-training largely govern the degree of diffuse redundancy and the "critical mass" of neurons needed often depends on the downstream task, suggesting that there is a task-inherent redundancy-performance Pareto frontier. Our findings shed light on the nature of representations learned by pre-trained deep neural networks and suggest that entire layers might not be necessary to perform many downstream tasks. We investigate the potential for exploiting this redundancy to achieve efficient generalization for downstream tasks and also draw caution to certain possible unintended consequences.

Text-To-Concept via Cross-Model Alignment

We observe that the mapping between an image's representation in one model to its representation in another can be learned surprisingly well with just a linear layer, even across diverse models. Building on this observation, we propose $\textit{text-to-concept}$, where features from a fixed pretrained model are aligned linearly to the CLIP space, so that text embeddings from CLIP's text encoder become directly comparable to the aligned features. With text-to-concept, we convert fixed off-the-shelf vision encoders to surprisingly strong zero-shot classifiers for free, with accuracy at times even surpassing that of CLIP, despite being much smaller models and trained on a small fraction of the data compared to CLIP. We show other immediate use-cases of text-to-concept, like building concept bottleneck models with no concept supervision, diagnosing distribution shifts in terms of human concepts, and retrieving images satisfying a set of text-based constraints. Lastly, we demonstrate the feasibility of $\textit{concept-to-text}$, where vectors in a model's feature space are decoded by first aligning to the CLIP before being fed to a GPT-based generative model. Our work suggests existing deep models, with presumably diverse architectures and training, represent input samples relatively similarly, and a two-way communication across model representation spaces and to humans (through language) is viable.

Strong Baselines for Parameter Efficient Few-Shot Fine-tuning

Few-shot classification (FSC) entails learning novel classes given only a few examples per class after a pre-training (or meta-training) phase on a set of base classes. Recent works have shown that simply fine-tuning a pre-trained Vision Transformer (ViT) on new test classes is a strong approach for FSC. Fine-tuning ViTs, however, is expensive in time, compute and storage. This has motivated the design of parameter efficient fine-tuning (PEFT) methods which fine-tune only a fraction of the Transformer's parameters. While these methods have shown promise, inconsistencies in experimental conditions make it difficult to disentangle their advantage from other experimental factors including the feature extractor architecture, pre-trained initialization and fine-tuning algorithm, amongst others. In our paper, we conduct a large-scale, experimentally consistent, empirical analysis to study PEFTs for few-shot image classification. Through a battery of over 1.8k controlled experiments on large-scale few-shot benchmarks including Meta-Dataset (MD) and ORBIT, we uncover novel insights on PEFTs that cast light on their efficacy in fine-tuning ViTs for few-shot classification. Through our controlled empirical study, we have two main findings: (i) Fine-tuning just the LayerNorm parameters (which we call LN-Tune) during few-shot adaptation is an extremely strong baseline across ViTs pre-trained with both self-supervised and supervised objectives, (ii) For self-supervised ViTs, we find that simply learning a set of scaling parameters for each attention matrix (which we call AttnScale) along with a domain-residual adapter (DRA) module leads to state-of-the-art performance (while being $\sim\!$ 9$\times$ more parameter-efficient) on MD. Our extensive empirical findings set strong baselines and call for rethinking the current design of PEFT methods for FSC.

Provable Robustness for Streaming Models with a Sliding Window

The literature on provable robustness in machine learning has primarily focused on static prediction problems, such as image classification, in which input samples are assumed to be independent and model performance is measured as an expectation over the input distribution. Robustness certificates are derived for individual input instances with the assumption that the model is evaluated on each instance separately. However, in many deep learning applications such as online content recommendation and stock market analysis, models use historical data to make predictions. Robustness certificates based on the assumption of independent input samples are not directly applicable in such scenarios. In this work, we focus on the provable robustness of machine learning models in the context of data streams, where inputs are presented as a sequence of potentially correlated items. We derive robustness certificates for models that use a fixed-size sliding window over the input stream. Our guarantees hold for the average model performance across the entire stream and are independent of stream size, making them suitable for large data streams. We perform experiments on speech detection and human activity recognition tasks and show that our certificates can produce meaningful performance guarantees against adversarial perturbations.

Adversarial Robustness of Learning-based Static Malware Classifiers

Malware detection has long been a stage for an ongoing arms race between malware authors and anti-virus systems. Solutions that utilize machine learning (ML) gain traction as the scale of this arms race increases. This trend, however, makes performing attacks directly on ML an attractive prospect for adversaries. We study this arms race from both perspectives in the context of MalConv, a popular convolutional neural network-based malware classifier that operates on raw bytes of files. First, we show that MalConv is vulnerable to adversarial patch attacks: appending a byte-level patch to malware files bypasses detection 94.3% of the time. Moreover, we develop a universal adversarial patch (UAP) attack where a single patch can drop the detection rate in constant time of any malware file that contains it by 80%. These patches are effective even being relatively small with respect to the original file size -- between 2%-8%. As a countermeasure, we then perform window ablation that allows us to apply de-randomized smoothing, a modern certified defense to patch attacks in vision tasks, to raw files. The resulting `smoothed-MalConv' can detect over 80% of malware that contains the universal patch and provides certified robustness up to 66%, outlining a promising step towards robust malware detection. To our knowledge, we are the first to apply universal adversarial patch attack and certified defense using ablations on byte level in the malware field.

Can AI-Generated Text be Reliably Detected?

The rapid progress of Large Language Models (LLMs) has made them capable of performing astonishingly well on various tasks including document completion and question answering. The unregulated use of these models, however, can potentially lead to malicious consequences such as plagiarism, generating fake news, spamming, etc. Therefore, reliable detection of AI-generated text can be critical to ensure the responsible use of LLMs. Recent works attempt to tackle this problem either using certain model signatures present in the generated text outputs or by applying watermarking techniques that imprint specific patterns onto them. In this paper, both empirically and theoretically, we show that these detectors are not reliable in practical scenarios. Empirically, we show that paraphrasing attacks, where a light paraphraser is applied on top of the generative text model, can break a whole range of detectors, including the ones using the watermarking schemes as well as neural network-based detectors and zero-shot classifiers. We then provide a theoretical impossibility result indicating that for a sufficiently good language model, even the best-possible detector can only perform marginally better than a random classifier. Finally, we show that even LLMs protected by watermarking schemes can be vulnerable against spoofing attacks where adversarial humans can infer hidden watermarking signatures and add them to their generated text to be detected as text generated by the LLMs, potentially causing reputational damages to their developers. We believe these results can open an honest conversation in the community regarding the ethical and reliable use of AI-generated text.

CUDA: Convolution-based Unlearnable Datasets

Large-scale training of modern deep learning models heavily relies on publicly available data on the web. This potentially unauthorized usage of online data leads to concerns regarding data privacy. Recent works aim to make unlearnable data for deep learning models by adding small, specially designed noises to tackle this issue. However, these methods are vulnerable to adversarial training (AT) and/or are computationally heavy. In this work, we propose a novel, model-free, Convolution-based Unlearnable DAtaset (CUDA) generation technique. CUDA is generated using controlled class-wise convolutions with filters that are randomly generated via a private key. CUDA encourages the network to learn the relation between filters and labels rather than informative features for classifying the clean data. We develop some theoretical analysis demonstrating that CUDA can successfully poison Gaussian mixture data by reducing the clean data performance of the optimal Bayes classifier. We also empirically demonstrate the effectiveness of CUDA with various datasets (CIFAR-10, CIFAR-100, ImageNet-100, and Tiny-ImageNet), and architectures (ResNet-18, VGG-16, Wide ResNet-34-10, DenseNet-121, DeIT, EfficientNetV2-S, and MobileNetV2). Our experiments show that CUDA is robust to various data augmentations and training approaches such as smoothing, AT with different budgets, transfer learning, and fine-tuning. For instance, training a ResNet-18 on ImageNet-100 CUDA achieves only 8.96$\%$, 40.08$\%$, and 20.58$\%$ clean test accuracies with empirical risk minimization (ERM), $L_{\infty}$ AT, and $L_{2}$ AT, respectively. Here, ERM on the clean training data achieves a clean test accuracy of 80.66$\%$. CUDA exhibits unlearnability effect with ERM even when only a fraction of the training dataset is perturbed. Furthermore, we also show that CUDA is robust to adaptive defenses designed specifically to break it.

Temporal Robustness against Data Poisoning

Data poisoning considers cases when an adversary maliciously inserts and removes training data to manipulate the behavior of machine learning algorithms. Traditional threat models of data poisoning center around a single metric, the number of poisoned samples. In consequence, existing defenses are essentially vulnerable in practice when poisoning more samples remains a feasible option for attackers. To address this issue, we leverage timestamps denoting the birth dates of data, which are often available but neglected in the past. Benefiting from these timestamps, we propose a temporal threat model of data poisoning and derive two novel metrics, earliness and duration, which respectively measure how long an attack started in advance and how long an attack lasted. With these metrics, we define the notions of temporal robustness against data poisoning, providing a meaningful sense of protection even with unbounded amounts of poisoned samples. We present a benchmark with an evaluation protocol simulating continuous data collection and periodic deployments of updated models, thus enabling empirical evaluation of temporal robustness. Lastly, we develop and also empirically verify a baseline defense, namely temporal aggregation, offering provable temporal robustness and highlighting the potential of our temporal modeling of data poisoning.