Dataset distillation offers a potential means to enhance data efficiency in deep learning. Recent studies have shown its ability to counteract backdoor risks present in original training samples. In this study, we delve into the theoretical aspects of backdoor attacks and dataset distillation based on kernel methods. We introduce two new theory-driven trigger pattern generation methods specialized for dataset distillation. Following a comprehensive set of analyses and experiments, we show that our optimization-based trigger design framework informs effective backdoor attacks on dataset distillation. Notably, datasets poisoned by our designed trigger prove resilient against conventional backdoor attack detection and mitigation methods. Our empirical results validate that the triggers developed using our approaches are proficient at executing resilient backdoor attacks.
Dataset distillation (DD) is a newly emerging research area aiming at alleviating the heavy computational load in training models on large datasets. It tries to distill a large dataset into a small and condensed one so that models trained on the distilled dataset can perform comparably with those trained on the full dataset when performing downstream tasks. Among the previous works in this area, there are three key problems that hinder the performance and availability of the existing DD methods: high time complexity, high space complexity, and low info-compactness. In this work, we simultaneously attempt to settle these three problems by moving the DD processes from conventionally used pixel space to latent space. Encoded by a pretrained generic autoencoder, latent codes in the latent space are naturally info-compact representations of the original images in much smaller sizes. After transferring three mainstream DD algorithms to latent space, we significantly reduce time and space consumption while achieving similar performance, allowing us to distill high-resolution datasets or target at greater data ratio that previous methods have failed. Besides, within the same storage budget, we can also quantitatively deliver more latent codes than pixel-level images, which further boosts the performance of our methods.
Dataset distillation reduces the storage and computational consumption of training a network by generating a small surrogate dataset that encapsulates rich information of the original large-scale one. However, previous distillation methods heavily rely on the sample-wise iterative optimization scheme. As the images-per-class (IPC) setting or image resolution grows larger, the necessary computation will demand overwhelming time and resources. In this work, we intend to incorporate generative diffusion techniques for computing the surrogate dataset. Observing that key factors for constructing an effective surrogate dataset are representativeness and diversity, we design additional minimax criteria in the generative training to enhance these facets for the generated images of diffusion models. We present a theoretical model of the process as hierarchical diffusion control demonstrating the flexibility of the diffusion process to target these criteria without jeopardizing the faithfulness of the sample to the desired distribution. The proposed method achieves state-of-the-art validation performance while demanding much less computational resources. Under the 100-IPC setting on ImageWoof, our method requires less than one-twentieth the distillation time of previous methods, yet yields even better performance. Source code and generated data are available in https://github.com/vimar-gu/MinimaxDiffusion.
Federated Unlearning (FU) aims to delete specific training data from an ML model trained using Federated Learning (FL). We introduce QuickDrop, an efficient and original FU method that utilizes dataset distillation (DD) to accelerate unlearning and drastically reduces computational overhead compared to existing approaches. In QuickDrop, each client uses DD to generate a compact dataset representative of the original training dataset, called a distilled dataset, and uses this compact dataset during unlearning. To unlearn specific knowledge from the global model, QuickDrop has clients execute Stochastic Gradient Ascent with samples from the distilled datasets, thus significantly reducing computational overhead compared to conventional FU methods. We further increase the efficiency of QuickDrop by ingeniously integrating DD into the FL training process. By reusing the gradient updates produced during FL training for DD, the overhead of creating distilled datasets becomes close to negligible. Evaluations on three standard datasets show that, with comparable accuracy guarantees, QuickDrop reduces the duration of unlearning by 463.8x compared to model retraining from scratch and 65.1x compared to existing FU approaches. We also demonstrate the scalability of QuickDrop with 100 clients and show its effectiveness while handling multiple unlearning operations.
Dataset distillation extracts a small set of synthetic training samples from a large dataset with the goal of achieving competitive performance on test data when trained on this sample. In this work, we tackle dataset distillation at its core by treating it directly as a bilevel optimization problem. Re-examining the foundational back-propagation through time method, we study the pronounced variance in the gradients, computational burden, and long-term dependencies. We introduce an improved method: Random Truncated Backpropagation Through Time (RaT-BPTT) to address them. RaT-BPTT incorporates a truncation coupled with a random window, effectively stabilizing the gradients and speeding up the optimization while covering long dependencies. This allows us to establish new state-of-the-art for a variety of standard dataset benchmarks. A deeper dive into the nature of distilled data unveils pronounced intercorrelation. In particular, subsets of distilled datasets tend to exhibit much worse performance than directly distilled smaller datasets of the same size. Leveraging RaT-BPTT, we devise a boosting mechanism that generates distilled datasets that contain subsets with near optimal performance across different data budgets.
Dataset distillation is a newly emerging task that synthesizes a small-size dataset used in training deep neural networks (DNNs) for reducing data storage and model training costs. The synthetic datasets are expected to capture the essence of the knowledge contained in real-world datasets such that the former yields a similar performance as the latter. Recent advancements in distillation methods have produced notable improvements in generating synthetic datasets. However, current state-of-the-art methods treat the entire synthetic dataset as a unified entity and optimize each synthetic instance equally. This static optimization approach may lead to performance degradation in dataset distillation. Specifically, we argue that static optimization can give rise to a coupling issue within the synthetic data, particularly when a larger amount of synthetic data is being optimized. This coupling issue, in turn, leads to the failure of the distilled dataset to extract the high-level features learned by the deep neural network (DNN) in the latter epochs. In this study, we propose a new dataset distillation strategy called Sequential Subset Matching (SeqMatch), which tackles this problem by adaptively optimizing the synthetic data to encourage sequential acquisition of knowledge during dataset distillation. Our analysis indicates that SeqMatch effectively addresses the coupling issue by sequentially generating the synthetic instances, thereby enhancing its performance significantly. Our proposed SeqMatch outperforms state-of-the-art methods in various datasets, including SVNH, CIFAR-10, CIFAR-100, and Tiny ImageNet. Our code is available at https://github.com/shqii1j/seqmatch.
In a backdoor attack, an adversary injects corrupted data into a model's training dataset in order to gain control over its predictions on images with a specific attacker-defined trigger. A typical corrupted training example requires altering both the image, by applying the trigger, and the label. Models trained on clean images, therefore, were considered safe from backdoor attacks. However, in some common machine learning scenarios, the training labels are provided by potentially malicious third-parties. This includes crowd-sourced annotation and knowledge distillation. We, hence, investigate a fundamental question: can we launch a successful backdoor attack by only corrupting labels? We introduce a novel approach to design label-only backdoor attacks, which we call FLIP, and demonstrate its strengths on three datasets (CIFAR-10, CIFAR-100, and Tiny-ImageNet) and four architectures (ResNet-32, ResNet-18, VGG-19, and Vision Transformer). With only 2% of CIFAR-10 labels corrupted, FLIP achieves a near-perfect attack success rate of 99.4% while suffering only a 1.8% drop in the clean test accuracy. Our approach builds upon the recent advances in trajectory matching, originally introduced for dataset distillation.
With the rise in interest of sparse neural networks, we study how neural network pruning with synthetic data leads to sparse networks with unique training properties. We find that distilled data, a synthetic summarization of the real data, paired with Iterative Magnitude Pruning (IMP) unveils a new class of sparse networks that are more stable to SGD noise on the real data, than either the dense model, or subnetworks found with real data in IMP. That is, synthetically chosen subnetworks often train to the same minima, or exhibit linear mode connectivity. We study this through linear interpolation, loss landscape visualizations, and measuring the diagonal of the hessian. While dataset distillation as a field is still young, we find that these properties lead to synthetic subnetworks matching the performance of traditional IMP with up to 150x less training points in settings where distilled data applies.
Dataset distillation plays a crucial role in creating compact datasets with similar training performance compared with original large-scale ones. This is essential for addressing the challenges of data storage and training costs. Prevalent methods facilitate knowledge transfer by matching the gradients, embedding distributions, or training trajectories of synthetic images with those of the sampled original images. Although there are various matching objectives, currently the strategy for selecting original images is limited to naive random sampling. We argue that random sampling overlooks the evenness of the selected sample distribution, which may result in noisy or biased matching targets. Besides, the sample diversity is also not constrained by random sampling. Additionally, current methods predominantly focus on single-dimensional matching, where information is not fully utilized. To address these challenges, we propose a novel matching strategy called Dataset Distillation by Bidirectional REpresentAtive Matching (DREAM+), which selects representative original images for bidirectional matching. DREAM+ is applicable to a variety of mainstream dataset distillation frameworks and significantly reduces the number of distillation iterations by more than 15 times without affecting performance. Given sufficient training time, DREAM+ can further improve the performance and achieve state-of-the-art results. We have released the code at github.com/NUS-HPC-AI-Lab/DREAM+.
Training large AI models typically requires large-scale datasets in the machine learning process, making training and parameter-tuning process both time-consuming and costly. Some researchers address this problem by carefully synthesizing a very small number of highly representative and informative samples from real-world datasets. This approach, known as Dataset Distillation (DD), proposes a perspective for data-efficient learning. Despite recent progress in this field, the performance of existing methods still cannot meet expectations, and distilled datasets cannot effectively replace original datasets. In this paper, unlike previous methods that focus solely on improving the effectiveness of student distillation, we recognize and leverage the important mutual influence between expert and student models. We observed that the smoothness of expert trajectories has a significant impact on subsequent student parameter alignment. Based on this, we propose an effective DD framework named AST, standing for Alignment with Smooth and high-quality expert Trajectories. We devise the integration of clipping loss and gradient penalty to regulate the rate of parameter changes in expert trajectory generation. To further refine the student parameter alignment with expert trajectory, we put forward representative initialization for the synthetic dataset and balanced inner-loop loss in response to the sensitivity exhibited towards randomly initialized variables during distillation. We also propose two enhancement strategies, namely intermediate matching loss and weight perturbation, to mitigate the potential occurrence of cumulative errors. We conduct extensive experiments on datasets of different scales, sizes, and resolutions. The results demonstrate that the proposed method significantly outperforms prior methods.
Dataset distillation aims to minimize the time and memory needed for training deep networks on large datasets, by creating a small set of synthetic images that has a similar generalization performance to that of the full dataset. However, current dataset distillation techniques fall short, showing a notable performance gap when compared to training on the original data. In this work, we are the first to argue that using just one synthetic subset for distillation will not yield optimal generalization performance. This is because the training dynamics of deep networks drastically change during the training. Hence, multiple synthetic subsets are required to capture the training dynamics at different phases of training. To address this issue, we propose Progressive Dataset Distillation (PDD). PDD synthesizes multiple small sets of synthetic images, each conditioned on the previous sets, and trains the model on the cumulative union of these subsets without requiring additional training time. Our extensive experiments show that PDD can effectively improve the performance of existing dataset distillation methods by up to 4.3%. In addition, our method for the first time enable generating considerably larger synthetic datasets.
Dataset distillation methods have achieved remarkable success in distilling a large dataset into a small set of representative samples. However, they are not designed to produce a distilled dataset that can be effectively used for facilitating self-supervised pre-training. To this end, we propose a novel problem of distilling an unlabeled dataset into a set of small synthetic samples for efficient self-supervised learning (SSL). We first prove that a gradient of synthetic samples with respect to a SSL objective in naive bilevel optimization is \textit{biased} due to the randomness originating from data augmentations or masking. To address this issue, we propose to minimize the mean squared error (MSE) between a model's representations of the synthetic examples and their corresponding learnable target feature representations for the inner objective, which does not introduce any randomness. Our primary motivation is that the model obtained by the proposed inner optimization can mimic the \textit{self-supervised target model}. To achieve this, we also introduce the MSE between representations of the inner model and the self-supervised target model on the original full dataset for outer optimization. Lastly, assuming that a feature extractor is fixed, we only optimize a linear head on top of the feature extractor, which allows us to reduce the computational cost and obtain a closed-form solution of the head with kernel ridge regression. We empirically validate the effectiveness of our method on various applications involving transfer learning.
The ultimate goal of Dataset Distillation is to synthesize a small synthetic dataset such that a model trained on this synthetic set will perform equally well as a model trained on the full, real dataset. Until now, no method of Dataset Distillation has reached this completely lossless goal, in part due to the fact that previous methods only remain effective when the total number of synthetic samples is extremely small. Since only so much information can be contained in such a small number of samples, it seems that to achieve truly loss dataset distillation, we must develop a distillation method that remains effective as the size of the synthetic dataset grows. In this work, we present such an algorithm and elucidate why existing methods fail to generate larger, high-quality synthetic sets. Current state-of-the-art methods rely on trajectory-matching, or optimizing the synthetic data to induce similar long-term training dynamics as the real data. We empirically find that the training stage of the trajectories we choose to match (i.e., early or late) greatly affects the effectiveness of the distilled dataset. Specifically, early trajectories (where the teacher network learns easy patterns) work well for a low-cardinality synthetic set since there are fewer examples wherein to distribute the necessary information. Conversely, late trajectories (where the teacher network learns hard patterns) provide better signals for larger synthetic sets since there are now enough samples to represent the necessary complex patterns. Based on our findings, we propose to align the difficulty of the generated patterns with the size of the synthetic dataset. In doing so, we successfully scale trajectory matching-based methods to larger synthetic datasets, achieving lossless dataset distillation for the very first time. Code and distilled datasets are available at https://gzyaftermath.github.io/DATM.
Dataset Distillation (DD) is a prominent technique that encapsulates knowledge from a large-scale original dataset into a small synthetic dataset for efficient training. Meanwhile, Pre-trained Models (PTMs) function as knowledge repositories, containing extensive information from the original dataset. This naturally raises a question: Can PTMs effectively transfer knowledge to synthetic datasets, guiding DD accurately? To this end, we conduct preliminary experiments, confirming the contribution of PTMs to DD. Afterwards, we systematically study different options in PTMs, including initialization parameters, model architecture, training epoch and domain knowledge, revealing that: 1) Increasing model diversity enhances the performance of synthetic datasets; 2) Sub-optimal models can also assist in DD and outperform well-trained ones in certain cases; 3) Domain-specific PTMs are not mandatory for DD, but a reasonable domain match is crucial. Finally, by selecting optimal options, we significantly improve the cross-architecture generalization over baseline DD methods. We hope our work will facilitate researchers to develop better DD techniques. Our code is available at https://github.com/yaolu-zjut/DDInterpreter.
Researchers have long tried to minimize training costs in deep learning while maintaining strong generalization across diverse datasets. Emerging research on dataset distillation aims to reduce training costs by creating a small synthetic set that contains the information of a larger real dataset and ultimately achieves test accuracy equivalent to a model trained on the whole dataset. Unfortunately, the synthetic data generated by previous methods are not guaranteed to distribute and discriminate as well as the original training data, and they incur significant computational costs. Despite promising results, there still exists a significant performance gap between models trained on condensed synthetic sets and those trained on the whole dataset. In this paper, we address these challenges using efficient Dataset Distillation with Attention Matching (DataDAM), achieving state-of-the-art performance while reducing training costs. Specifically, we learn synthetic images by matching the spatial attention maps of real and synthetic data generated by different layers within a family of randomly initialized neural networks. Our method outperforms the prior methods on several datasets, including CIFAR10/100, TinyImageNet, ImageNet-1K, and subsets of ImageNet-1K across most of the settings, and achieves improvements of up to 6.5% and 4.1% on CIFAR100 and ImageNet-1K, respectively. We also show that our high-quality distilled images have practical benefits for downstream applications, such as continual learning and neural architecture search.
Federated training of Graph Neural Networks (GNN) has become popular in recent years due to its ability to perform graph-related tasks under data isolation scenarios while preserving data privacy. However, graph heterogeneity issues in federated GNN systems continue to pose challenges. Existing frameworks address the problem by representing local tasks using different statistics and relating them through a simple aggregation mechanism. However, these approaches suffer from limited efficiency from two aspects: low quality of task-relatedness quantification and inefficacy of exploiting the collaboration structure. To address these issues, we propose FedGKD, a novel federated GNN framework that utilizes a novel client-side graph dataset distillation method to extract task features that better describe task-relatedness, and introduces a novel server-side aggregation mechanism that is aware of the global collaboration structure. We conduct extensive experiments on six real-world datasets of different scales, demonstrating our framework's outperformance.
In this paper, we consider the intersection of two problems in machine learning: Multi-Source Domain Adaptation (MSDA) and Dataset Distillation (DD). On the one hand, the first considers adapting multiple heterogeneous labeled source domains to an unlabeled target domain. On the other hand, the second attacks the problem of synthesizing a small summary containing all the information about the datasets. We thus consider a new problem called MSDA-DD. To solve it, we adapt previous works in the MSDA literature, such as Wasserstein Barycenter Transport and Dataset Dictionary Learning, as well as DD method Distribution Matching. We thoroughly experiment with this novel problem on four benchmarks (Caltech-Office 10, Tennessee-Eastman Process, Continuous Stirred Tank Reactor, and Case Western Reserve University), where we show that, even with as little as 1 sample per class, one achieves state-of-the-art adaptation performance.
Dataset distillation methods have demonstrated remarkable performance for neural networks trained with very limited training data. However, a significant challenge arises in the form of \textit{architecture overfitting}: the distilled training dataset synthesized by a specific network architecture (i.e., training network) generates poor performance when trained by other network architectures (i.e., test networks), especially when the test networks have a larger capacity than the training network. This paper introduces a series of approaches to mitigate this issue. Among them, DropPath renders the large model to be an implicit ensemble of its sub-networks, and knowledge distillation ensures each sub-network acts similarly to the small but well-performing teacher network. These methods, characterized by their smoothing effects, significantly mitigate architecture overfitting. We conduct extensive experiments to demonstrate the effectiveness and generality of our methods. Particularly, across various scenarios involving different tasks and different sizes of distilled data, our approaches significantly mitigate architecture overfitting. Furthermore, our approaches achieve comparable or even superior performance when the test network is larger than the training network.
State-of-the-art deep neural networks are trained with large amounts (millions or even billions) of data. The expensive computation and memory costs make it difficult to train them on limited hardware resources, especially for recent popular large language models (LLM) and computer vision models (CV). Recent popular dataset distillation methods are thus developed, aiming to reduce the number of training samples via synthesizing small-scale datasets via gradient matching. However, as the gradient calculation is coupled with the specific network architecture, the synthesized dataset is biased and performs poorly when used for training unseen architectures. To address these limitations, we present dataset quantization (DQ), a new framework to compress large-scale datasets into small subsets which can be used for training any neural network architectures. Extensive experiments demonstrate that DQ is able to generate condensed small datasets for training unseen network architectures with state-of-the-art compression ratios for lossless model training. To the best of our knowledge, DQ is the first method that can successfully distill large-scale datasets such as ImageNet-1k with a state-of-the-art compression ratio. Notably, with 60% data from ImageNet and 20% data from Alpaca's instruction tuning data, the models can be trained with negligible or no performance drop for both vision tasks (including classification, semantic segmentation, and object detection) as well as language tasks (including instruction tuning tasks such as BBH and DROP).
Dataset distillation methods reduce large-scale datasets to smaller sets of synthetic data, preserving sufficient information to quickly train a new model from scratch. However, prior work on dataset distillation has focused exclusively on image classification datasets, whereas modern large-scale datasets are primarily vision-language datasets. In this work, we design the first vision-language dataset distillation method, building on the idea of trajectory matching. A key challenge is that vision-language datasets do not have a set of discrete classes. To overcome this, our proposed method jointly distills image-text pairs in a contrastive formulation. Further, we leverage Low-Rank Adaptation (LoRA) matching to enable more efficient and effective trajectory matching in complex modern vision-language models. Since there are no existing baselines, we compare our distillation approach with three adapted vision-language coreset selection methods. We demonstrate significant improvements on the challenging Flickr30K and COCO retrieval benchmarks: for example, on Flickr30K, the best coreset selection method selecting 1000 image-text pairs for training achieves only 5.6% image-to-text retrieval accuracy (i.e., recall@1); in contrast, our dataset distillation almost doubles that to 9.9% with just 100 training pairs, an order of magnitude fewer.